1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 37 */ 38 39 /* 40 * External virtual filesystem routines 41 */ 42 43 #include <sys/cdefs.h> 44 __FBSDID("$FreeBSD$"); 45 46 #include "opt_ddb.h" 47 #include "opt_watchdog.h" 48 49 #include <sys/param.h> 50 #include <sys/systm.h> 51 #include <sys/bio.h> 52 #include <sys/buf.h> 53 #include <sys/capsicum.h> 54 #include <sys/condvar.h> 55 #include <sys/conf.h> 56 #include <sys/counter.h> 57 #include <sys/dirent.h> 58 #include <sys/event.h> 59 #include <sys/eventhandler.h> 60 #include <sys/extattr.h> 61 #include <sys/file.h> 62 #include <sys/fcntl.h> 63 #include <sys/jail.h> 64 #include <sys/kdb.h> 65 #include <sys/kernel.h> 66 #include <sys/kthread.h> 67 #include <sys/ktr.h> 68 #include <sys/lockf.h> 69 #include <sys/malloc.h> 70 #include <sys/mount.h> 71 #include <sys/namei.h> 72 #include <sys/pctrie.h> 73 #include <sys/priv.h> 74 #include <sys/reboot.h> 75 #include <sys/refcount.h> 76 #include <sys/rwlock.h> 77 #include <sys/sched.h> 78 #include <sys/sleepqueue.h> 79 #include <sys/smp.h> 80 #include <sys/stat.h> 81 #include <sys/sysctl.h> 82 #include <sys/syslog.h> 83 #include <sys/vmmeter.h> 84 #include <sys/vnode.h> 85 #include <sys/watchdog.h> 86 87 #include <machine/stdarg.h> 88 89 #include <security/mac/mac_framework.h> 90 91 #include <vm/vm.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_extern.h> 94 #include <vm/pmap.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_page.h> 97 #include <vm/vm_kern.h> 98 #include <vm/uma.h> 99 100 #ifdef DDB 101 #include <ddb/ddb.h> 102 #endif 103 104 static void delmntque(struct vnode *vp); 105 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 106 int slpflag, int slptimeo); 107 static void syncer_shutdown(void *arg, int howto); 108 static int vtryrecycle(struct vnode *vp); 109 static void v_init_counters(struct vnode *); 110 static void v_incr_devcount(struct vnode *); 111 static void v_decr_devcount(struct vnode *); 112 static void vgonel(struct vnode *); 113 static void vfs_knllock(void *arg); 114 static void vfs_knlunlock(void *arg); 115 static void vfs_knl_assert_locked(void *arg); 116 static void vfs_knl_assert_unlocked(void *arg); 117 static void vnlru_return_batches(struct vfsops *mnt_op); 118 static void destroy_vpollinfo(struct vpollinfo *vi); 119 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 120 daddr_t startlbn, daddr_t endlbn); 121 static void vnlru_recalc(void); 122 123 /* 124 * These fences are intended for cases where some synchronization is 125 * needed between access of v_iflags and lockless vnode refcount (v_holdcnt 126 * and v_usecount) updates. Access to v_iflags is generally synchronized 127 * by the interlock, but we have some internal assertions that check vnode 128 * flags without acquiring the lock. Thus, these fences are INVARIANTS-only 129 * for now. 130 */ 131 #ifdef INVARIANTS 132 #define VNODE_REFCOUNT_FENCE_ACQ() atomic_thread_fence_acq() 133 #define VNODE_REFCOUNT_FENCE_REL() atomic_thread_fence_rel() 134 #else 135 #define VNODE_REFCOUNT_FENCE_ACQ() 136 #define VNODE_REFCOUNT_FENCE_REL() 137 #endif 138 139 /* 140 * Number of vnodes in existence. Increased whenever getnewvnode() 141 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode. 142 */ 143 static u_long __exclusive_cache_line numvnodes; 144 145 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 146 "Number of vnodes in existence"); 147 148 static counter_u64_t vnodes_created; 149 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, 150 "Number of vnodes created by getnewvnode"); 151 152 static u_long mnt_free_list_batch = 128; 153 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW, 154 &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list"); 155 156 /* 157 * Conversion tables for conversion from vnode types to inode formats 158 * and back. 159 */ 160 enum vtype iftovt_tab[16] = { 161 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 162 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON 163 }; 164 int vttoif_tab[10] = { 165 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 166 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 167 }; 168 169 /* 170 * List of vnodes that are ready for recycling. 171 */ 172 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 173 174 /* 175 * "Free" vnode target. Free vnodes are rarely completely free, but are 176 * just ones that are cheap to recycle. Usually they are for files which 177 * have been stat'd but not read; these usually have inode and namecache 178 * data attached to them. This target is the preferred minimum size of a 179 * sub-cache consisting mostly of such files. The system balances the size 180 * of this sub-cache with its complement to try to prevent either from 181 * thrashing while the other is relatively inactive. The targets express 182 * a preference for the best balance. 183 * 184 * "Above" this target there are 2 further targets (watermarks) related 185 * to recyling of free vnodes. In the best-operating case, the cache is 186 * exactly full, the free list has size between vlowat and vhiwat above the 187 * free target, and recycling from it and normal use maintains this state. 188 * Sometimes the free list is below vlowat or even empty, but this state 189 * is even better for immediate use provided the cache is not full. 190 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free 191 * ones) to reach one of these states. The watermarks are currently hard- 192 * coded as 4% and 9% of the available space higher. These and the default 193 * of 25% for wantfreevnodes are too large if the memory size is large. 194 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim 195 * whenever vnlru_proc() becomes active. 196 */ 197 static u_long wantfreevnodes; 198 static u_long freevnodes; 199 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, 200 &freevnodes, 0, "Number of \"free\" vnodes"); 201 202 static counter_u64_t recycles_count; 203 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 204 "Number of vnodes recycled to meet vnode cache targets"); 205 206 static counter_u64_t recycles_free_count; 207 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count, 208 "Number of free vnodes recycled to meet vnode cache targets"); 209 210 /* 211 * Various variables used for debugging the new implementation of 212 * reassignbuf(). 213 * XXX these are probably of (very) limited utility now. 214 */ 215 static int reassignbufcalls; 216 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS, 217 &reassignbufcalls, 0, "Number of calls to reassignbuf"); 218 219 static counter_u64_t deferred_inact; 220 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact, 221 "Number of times inactive processing was deferred"); 222 223 /* To keep more than one thread at a time from running vfs_getnewfsid */ 224 static struct mtx mntid_mtx; 225 226 /* 227 * Lock for any access to the following: 228 * vnode_free_list 229 * numvnodes 230 * freevnodes 231 */ 232 static struct mtx __exclusive_cache_line vnode_free_list_mtx; 233 234 /* Publicly exported FS */ 235 struct nfs_public nfs_pub; 236 237 static uma_zone_t buf_trie_zone; 238 239 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 240 static uma_zone_t vnode_zone; 241 static uma_zone_t vnodepoll_zone; 242 243 /* 244 * The workitem queue. 245 * 246 * It is useful to delay writes of file data and filesystem metadata 247 * for tens of seconds so that quickly created and deleted files need 248 * not waste disk bandwidth being created and removed. To realize this, 249 * we append vnodes to a "workitem" queue. When running with a soft 250 * updates implementation, most pending metadata dependencies should 251 * not wait for more than a few seconds. Thus, mounted on block devices 252 * are delayed only about a half the time that file data is delayed. 253 * Similarly, directory updates are more critical, so are only delayed 254 * about a third the time that file data is delayed. Thus, there are 255 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 256 * one each second (driven off the filesystem syncer process). The 257 * syncer_delayno variable indicates the next queue that is to be processed. 258 * Items that need to be processed soon are placed in this queue: 259 * 260 * syncer_workitem_pending[syncer_delayno] 261 * 262 * A delay of fifteen seconds is done by placing the request fifteen 263 * entries later in the queue: 264 * 265 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 266 * 267 */ 268 static int syncer_delayno; 269 static long syncer_mask; 270 LIST_HEAD(synclist, bufobj); 271 static struct synclist *syncer_workitem_pending; 272 /* 273 * The sync_mtx protects: 274 * bo->bo_synclist 275 * sync_vnode_count 276 * syncer_delayno 277 * syncer_state 278 * syncer_workitem_pending 279 * syncer_worklist_len 280 * rushjob 281 */ 282 static struct mtx sync_mtx; 283 static struct cv sync_wakeup; 284 285 #define SYNCER_MAXDELAY 32 286 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 287 static int syncdelay = 30; /* max time to delay syncing data */ 288 static int filedelay = 30; /* time to delay syncing files */ 289 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, 290 "Time to delay syncing files (in seconds)"); 291 static int dirdelay = 29; /* time to delay syncing directories */ 292 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, 293 "Time to delay syncing directories (in seconds)"); 294 static int metadelay = 28; /* time to delay syncing metadata */ 295 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, 296 "Time to delay syncing metadata (in seconds)"); 297 static int rushjob; /* number of slots to run ASAP */ 298 static int stat_rush_requests; /* number of times I/O speeded up */ 299 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, 300 "Number of times I/O speeded up (rush requests)"); 301 302 /* 303 * When shutting down the syncer, run it at four times normal speed. 304 */ 305 #define SYNCER_SHUTDOWN_SPEEDUP 4 306 static int sync_vnode_count; 307 static int syncer_worklist_len; 308 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 309 syncer_state; 310 311 /* Target for maximum number of vnodes. */ 312 u_long desiredvnodes; 313 static u_long gapvnodes; /* gap between wanted and desired */ 314 static u_long vhiwat; /* enough extras after expansion */ 315 static u_long vlowat; /* minimal extras before expansion */ 316 static u_long vstir; /* nonzero to stir non-free vnodes */ 317 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ 318 319 /* 320 * Note that no attempt is made to sanitize these parameters. 321 */ 322 static int 323 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS) 324 { 325 u_long val; 326 int error; 327 328 val = desiredvnodes; 329 error = sysctl_handle_long(oidp, &val, 0, req); 330 if (error != 0 || req->newptr == NULL) 331 return (error); 332 333 if (val == desiredvnodes) 334 return (0); 335 mtx_lock(&vnode_free_list_mtx); 336 desiredvnodes = val; 337 wantfreevnodes = desiredvnodes / 4; 338 vnlru_recalc(); 339 mtx_unlock(&vnode_free_list_mtx); 340 /* 341 * XXX There is no protection against multiple threads changing 342 * desiredvnodes at the same time. Locking above only helps vnlru and 343 * getnewvnode. 344 */ 345 vfs_hash_changesize(desiredvnodes); 346 cache_changesize(desiredvnodes); 347 return (0); 348 } 349 350 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, 351 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, 352 "UL", "Target for maximum number of vnodes"); 353 354 static int 355 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS) 356 { 357 u_long val; 358 int error; 359 360 val = wantfreevnodes; 361 error = sysctl_handle_long(oidp, &val, 0, req); 362 if (error != 0 || req->newptr == NULL) 363 return (error); 364 365 if (val == wantfreevnodes) 366 return (0); 367 mtx_lock(&vnode_free_list_mtx); 368 wantfreevnodes = val; 369 vnlru_recalc(); 370 mtx_unlock(&vnode_free_list_mtx); 371 return (0); 372 } 373 374 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes, 375 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, 376 "UL", "Target for minimum number of \"free\" vnodes"); 377 378 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 379 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)"); 380 static int vnlru_nowhere; 381 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 382 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 383 384 static int 385 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS) 386 { 387 struct vnode *vp; 388 struct nameidata nd; 389 char *buf; 390 unsigned long ndflags; 391 int error; 392 393 if (req->newptr == NULL) 394 return (EINVAL); 395 if (req->newlen >= PATH_MAX) 396 return (E2BIG); 397 398 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK); 399 error = SYSCTL_IN(req, buf, req->newlen); 400 if (error != 0) 401 goto out; 402 403 buf[req->newlen] = '\0'; 404 405 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME; 406 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread); 407 if ((error = namei(&nd)) != 0) 408 goto out; 409 vp = nd.ni_vp; 410 411 if (VN_IS_DOOMED(vp)) { 412 /* 413 * This vnode is being recycled. Return != 0 to let the caller 414 * know that the sysctl had no effect. Return EAGAIN because a 415 * subsequent call will likely succeed (since namei will create 416 * a new vnode if necessary) 417 */ 418 error = EAGAIN; 419 goto putvnode; 420 } 421 422 counter_u64_add(recycles_count, 1); 423 vgone(vp); 424 putvnode: 425 NDFREE(&nd, 0); 426 out: 427 free(buf, M_TEMP); 428 return (error); 429 } 430 431 static int 432 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS) 433 { 434 struct thread *td = curthread; 435 struct vnode *vp; 436 struct file *fp; 437 int error; 438 int fd; 439 440 if (req->newptr == NULL) 441 return (EBADF); 442 443 error = sysctl_handle_int(oidp, &fd, 0, req); 444 if (error != 0) 445 return (error); 446 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp); 447 if (error != 0) 448 return (error); 449 vp = fp->f_vnode; 450 451 error = vn_lock(vp, LK_EXCLUSIVE); 452 if (error != 0) 453 goto drop; 454 455 counter_u64_add(recycles_count, 1); 456 vgone(vp); 457 VOP_UNLOCK(vp); 458 drop: 459 fdrop(fp, td); 460 return (error); 461 } 462 463 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode, 464 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 465 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname"); 466 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode, 467 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 468 sysctl_ftry_reclaim_vnode, "I", 469 "Try to reclaim a vnode by its file descriptor"); 470 471 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 472 static int vnsz2log; 473 474 /* 475 * Support for the bufobj clean & dirty pctrie. 476 */ 477 static void * 478 buf_trie_alloc(struct pctrie *ptree) 479 { 480 481 return uma_zalloc(buf_trie_zone, M_NOWAIT); 482 } 483 484 static void 485 buf_trie_free(struct pctrie *ptree, void *node) 486 { 487 488 uma_zfree(buf_trie_zone, node); 489 } 490 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free); 491 492 /* 493 * Initialize the vnode management data structures. 494 * 495 * Reevaluate the following cap on the number of vnodes after the physical 496 * memory size exceeds 512GB. In the limit, as the physical memory size 497 * grows, the ratio of the memory size in KB to vnodes approaches 64:1. 498 */ 499 #ifndef MAXVNODES_MAX 500 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */ 501 #endif 502 503 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 504 505 static struct vnode * 506 vn_alloc_marker(struct mount *mp) 507 { 508 struct vnode *vp; 509 510 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 511 vp->v_type = VMARKER; 512 vp->v_mount = mp; 513 514 return (vp); 515 } 516 517 static void 518 vn_free_marker(struct vnode *vp) 519 { 520 521 MPASS(vp->v_type == VMARKER); 522 free(vp, M_VNODE_MARKER); 523 } 524 525 /* 526 * Initialize a vnode as it first enters the zone. 527 */ 528 static int 529 vnode_init(void *mem, int size, int flags) 530 { 531 struct vnode *vp; 532 533 vp = mem; 534 bzero(vp, size); 535 /* 536 * Setup locks. 537 */ 538 vp->v_vnlock = &vp->v_lock; 539 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 540 /* 541 * By default, don't allow shared locks unless filesystems opt-in. 542 */ 543 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, 544 LK_NOSHARE | LK_IS_VNODE); 545 /* 546 * Initialize bufobj. 547 */ 548 bufobj_init(&vp->v_bufobj, vp); 549 /* 550 * Initialize namecache. 551 */ 552 LIST_INIT(&vp->v_cache_src); 553 TAILQ_INIT(&vp->v_cache_dst); 554 /* 555 * Initialize rangelocks. 556 */ 557 rangelock_init(&vp->v_rl); 558 return (0); 559 } 560 561 /* 562 * Free a vnode when it is cleared from the zone. 563 */ 564 static void 565 vnode_fini(void *mem, int size) 566 { 567 struct vnode *vp; 568 struct bufobj *bo; 569 570 vp = mem; 571 rangelock_destroy(&vp->v_rl); 572 lockdestroy(vp->v_vnlock); 573 mtx_destroy(&vp->v_interlock); 574 bo = &vp->v_bufobj; 575 rw_destroy(BO_LOCKPTR(bo)); 576 } 577 578 /* 579 * Provide the size of NFS nclnode and NFS fh for calculation of the 580 * vnode memory consumption. The size is specified directly to 581 * eliminate dependency on NFS-private header. 582 * 583 * Other filesystems may use bigger or smaller (like UFS and ZFS) 584 * private inode data, but the NFS-based estimation is ample enough. 585 * Still, we care about differences in the size between 64- and 32-bit 586 * platforms. 587 * 588 * Namecache structure size is heuristically 589 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. 590 */ 591 #ifdef _LP64 592 #define NFS_NCLNODE_SZ (528 + 64) 593 #define NC_SZ 148 594 #else 595 #define NFS_NCLNODE_SZ (360 + 32) 596 #define NC_SZ 92 597 #endif 598 599 static void 600 vntblinit(void *dummy __unused) 601 { 602 u_int i; 603 int physvnodes, virtvnodes; 604 605 /* 606 * Desiredvnodes is a function of the physical memory size and the 607 * kernel's heap size. Generally speaking, it scales with the 608 * physical memory size. The ratio of desiredvnodes to the physical 609 * memory size is 1:16 until desiredvnodes exceeds 98,304. 610 * Thereafter, the 611 * marginal ratio of desiredvnodes to the physical memory size is 612 * 1:64. However, desiredvnodes is limited by the kernel's heap 613 * size. The memory required by desiredvnodes vnodes and vm objects 614 * must not exceed 1/10th of the kernel's heap size. 615 */ 616 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 617 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; 618 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + 619 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); 620 desiredvnodes = min(physvnodes, virtvnodes); 621 if (desiredvnodes > MAXVNODES_MAX) { 622 if (bootverbose) 623 printf("Reducing kern.maxvnodes %lu -> %lu\n", 624 desiredvnodes, MAXVNODES_MAX); 625 desiredvnodes = MAXVNODES_MAX; 626 } 627 wantfreevnodes = desiredvnodes / 4; 628 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 629 TAILQ_INIT(&vnode_free_list); 630 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 631 /* 632 * The lock is taken to appease WITNESS. 633 */ 634 mtx_lock(&vnode_free_list_mtx); 635 vnlru_recalc(); 636 mtx_unlock(&vnode_free_list_mtx); 637 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 638 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0); 639 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 640 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 641 /* 642 * Preallocate enough nodes to support one-per buf so that 643 * we can not fail an insert. reassignbuf() callers can not 644 * tolerate the insertion failure. 645 */ 646 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), 647 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 648 UMA_ZONE_NOFREE | UMA_ZONE_VM); 649 uma_prealloc(buf_trie_zone, nbuf); 650 651 vnodes_created = counter_u64_alloc(M_WAITOK); 652 recycles_count = counter_u64_alloc(M_WAITOK); 653 recycles_free_count = counter_u64_alloc(M_WAITOK); 654 deferred_inact = counter_u64_alloc(M_WAITOK); 655 656 /* 657 * Initialize the filesystem syncer. 658 */ 659 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 660 &syncer_mask); 661 syncer_maxdelay = syncer_mask + 1; 662 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 663 cv_init(&sync_wakeup, "syncer"); 664 for (i = 1; i <= sizeof(struct vnode); i <<= 1) 665 vnsz2log++; 666 vnsz2log--; 667 } 668 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 669 670 671 /* 672 * Mark a mount point as busy. Used to synchronize access and to delay 673 * unmounting. Eventually, mountlist_mtx is not released on failure. 674 * 675 * vfs_busy() is a custom lock, it can block the caller. 676 * vfs_busy() only sleeps if the unmount is active on the mount point. 677 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 678 * vnode belonging to mp. 679 * 680 * Lookup uses vfs_busy() to traverse mount points. 681 * root fs var fs 682 * / vnode lock A / vnode lock (/var) D 683 * /var vnode lock B /log vnode lock(/var/log) E 684 * vfs_busy lock C vfs_busy lock F 685 * 686 * Within each file system, the lock order is C->A->B and F->D->E. 687 * 688 * When traversing across mounts, the system follows that lock order: 689 * 690 * C->A->B 691 * | 692 * +->F->D->E 693 * 694 * The lookup() process for namei("/var") illustrates the process: 695 * VOP_LOOKUP() obtains B while A is held 696 * vfs_busy() obtains a shared lock on F while A and B are held 697 * vput() releases lock on B 698 * vput() releases lock on A 699 * VFS_ROOT() obtains lock on D while shared lock on F is held 700 * vfs_unbusy() releases shared lock on F 701 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 702 * Attempt to lock A (instead of vp_crossmp) while D is held would 703 * violate the global order, causing deadlocks. 704 * 705 * dounmount() locks B while F is drained. 706 */ 707 int 708 vfs_busy(struct mount *mp, int flags) 709 { 710 711 MPASS((flags & ~MBF_MASK) == 0); 712 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 713 714 if (vfs_op_thread_enter(mp)) { 715 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 716 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0); 717 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0); 718 vfs_mp_count_add_pcpu(mp, ref, 1); 719 vfs_mp_count_add_pcpu(mp, lockref, 1); 720 vfs_op_thread_exit(mp); 721 if (flags & MBF_MNTLSTLOCK) 722 mtx_unlock(&mountlist_mtx); 723 return (0); 724 } 725 726 MNT_ILOCK(mp); 727 vfs_assert_mount_counters(mp); 728 MNT_REF(mp); 729 /* 730 * If mount point is currently being unmounted, sleep until the 731 * mount point fate is decided. If thread doing the unmounting fails, 732 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 733 * that this mount point has survived the unmount attempt and vfs_busy 734 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 735 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 736 * about to be really destroyed. vfs_busy needs to release its 737 * reference on the mount point in this case and return with ENOENT, 738 * telling the caller that mount mount it tried to busy is no longer 739 * valid. 740 */ 741 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 742 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 743 MNT_REL(mp); 744 MNT_IUNLOCK(mp); 745 CTR1(KTR_VFS, "%s: failed busying before sleeping", 746 __func__); 747 return (ENOENT); 748 } 749 if (flags & MBF_MNTLSTLOCK) 750 mtx_unlock(&mountlist_mtx); 751 mp->mnt_kern_flag |= MNTK_MWAIT; 752 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 753 if (flags & MBF_MNTLSTLOCK) 754 mtx_lock(&mountlist_mtx); 755 MNT_ILOCK(mp); 756 } 757 if (flags & MBF_MNTLSTLOCK) 758 mtx_unlock(&mountlist_mtx); 759 mp->mnt_lockref++; 760 MNT_IUNLOCK(mp); 761 return (0); 762 } 763 764 /* 765 * Free a busy filesystem. 766 */ 767 void 768 vfs_unbusy(struct mount *mp) 769 { 770 int c; 771 772 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 773 774 if (vfs_op_thread_enter(mp)) { 775 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 776 vfs_mp_count_sub_pcpu(mp, lockref, 1); 777 vfs_mp_count_sub_pcpu(mp, ref, 1); 778 vfs_op_thread_exit(mp); 779 return; 780 } 781 782 MNT_ILOCK(mp); 783 vfs_assert_mount_counters(mp); 784 MNT_REL(mp); 785 c = --mp->mnt_lockref; 786 if (mp->mnt_vfs_ops == 0) { 787 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 788 MNT_IUNLOCK(mp); 789 return; 790 } 791 if (c < 0) 792 vfs_dump_mount_counters(mp); 793 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 794 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 795 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 796 mp->mnt_kern_flag &= ~MNTK_DRAINING; 797 wakeup(&mp->mnt_lockref); 798 } 799 MNT_IUNLOCK(mp); 800 } 801 802 /* 803 * Lookup a mount point by filesystem identifier. 804 */ 805 struct mount * 806 vfs_getvfs(fsid_t *fsid) 807 { 808 struct mount *mp; 809 810 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 811 mtx_lock(&mountlist_mtx); 812 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 813 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 814 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 815 vfs_ref(mp); 816 mtx_unlock(&mountlist_mtx); 817 return (mp); 818 } 819 } 820 mtx_unlock(&mountlist_mtx); 821 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 822 return ((struct mount *) 0); 823 } 824 825 /* 826 * Lookup a mount point by filesystem identifier, busying it before 827 * returning. 828 * 829 * To avoid congestion on mountlist_mtx, implement simple direct-mapped 830 * cache for popular filesystem identifiers. The cache is lockess, using 831 * the fact that struct mount's are never freed. In worst case we may 832 * get pointer to unmounted or even different filesystem, so we have to 833 * check what we got, and go slow way if so. 834 */ 835 struct mount * 836 vfs_busyfs(fsid_t *fsid) 837 { 838 #define FSID_CACHE_SIZE 256 839 typedef struct mount * volatile vmp_t; 840 static vmp_t cache[FSID_CACHE_SIZE]; 841 struct mount *mp; 842 int error; 843 uint32_t hash; 844 845 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 846 hash = fsid->val[0] ^ fsid->val[1]; 847 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); 848 mp = cache[hash]; 849 if (mp == NULL || 850 mp->mnt_stat.f_fsid.val[0] != fsid->val[0] || 851 mp->mnt_stat.f_fsid.val[1] != fsid->val[1]) 852 goto slow; 853 if (vfs_busy(mp, 0) != 0) { 854 cache[hash] = NULL; 855 goto slow; 856 } 857 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 858 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) 859 return (mp); 860 else 861 vfs_unbusy(mp); 862 863 slow: 864 mtx_lock(&mountlist_mtx); 865 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 866 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 867 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 868 error = vfs_busy(mp, MBF_MNTLSTLOCK); 869 if (error) { 870 cache[hash] = NULL; 871 mtx_unlock(&mountlist_mtx); 872 return (NULL); 873 } 874 cache[hash] = mp; 875 return (mp); 876 } 877 } 878 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 879 mtx_unlock(&mountlist_mtx); 880 return ((struct mount *) 0); 881 } 882 883 /* 884 * Check if a user can access privileged mount options. 885 */ 886 int 887 vfs_suser(struct mount *mp, struct thread *td) 888 { 889 int error; 890 891 if (jailed(td->td_ucred)) { 892 /* 893 * If the jail of the calling thread lacks permission for 894 * this type of file system, deny immediately. 895 */ 896 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag)) 897 return (EPERM); 898 899 /* 900 * If the file system was mounted outside the jail of the 901 * calling thread, deny immediately. 902 */ 903 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 904 return (EPERM); 905 } 906 907 /* 908 * If file system supports delegated administration, we don't check 909 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 910 * by the file system itself. 911 * If this is not the user that did original mount, we check for 912 * the PRIV_VFS_MOUNT_OWNER privilege. 913 */ 914 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 915 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 916 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 917 return (error); 918 } 919 return (0); 920 } 921 922 /* 923 * Get a new unique fsid. Try to make its val[0] unique, since this value 924 * will be used to create fake device numbers for stat(). Also try (but 925 * not so hard) make its val[0] unique mod 2^16, since some emulators only 926 * support 16-bit device numbers. We end up with unique val[0]'s for the 927 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 928 * 929 * Keep in mind that several mounts may be running in parallel. Starting 930 * the search one past where the previous search terminated is both a 931 * micro-optimization and a defense against returning the same fsid to 932 * different mounts. 933 */ 934 void 935 vfs_getnewfsid(struct mount *mp) 936 { 937 static uint16_t mntid_base; 938 struct mount *nmp; 939 fsid_t tfsid; 940 int mtype; 941 942 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 943 mtx_lock(&mntid_mtx); 944 mtype = mp->mnt_vfc->vfc_typenum; 945 tfsid.val[1] = mtype; 946 mtype = (mtype & 0xFF) << 24; 947 for (;;) { 948 tfsid.val[0] = makedev(255, 949 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 950 mntid_base++; 951 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 952 break; 953 vfs_rel(nmp); 954 } 955 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 956 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 957 mtx_unlock(&mntid_mtx); 958 } 959 960 /* 961 * Knob to control the precision of file timestamps: 962 * 963 * 0 = seconds only; nanoseconds zeroed. 964 * 1 = seconds and nanoseconds, accurate within 1/HZ. 965 * 2 = seconds and nanoseconds, truncated to microseconds. 966 * >=3 = seconds and nanoseconds, maximum precision. 967 */ 968 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 969 970 static int timestamp_precision = TSP_USEC; 971 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 972 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 973 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " 974 "3+: sec + ns (max. precision))"); 975 976 /* 977 * Get a current timestamp. 978 */ 979 void 980 vfs_timestamp(struct timespec *tsp) 981 { 982 struct timeval tv; 983 984 switch (timestamp_precision) { 985 case TSP_SEC: 986 tsp->tv_sec = time_second; 987 tsp->tv_nsec = 0; 988 break; 989 case TSP_HZ: 990 getnanotime(tsp); 991 break; 992 case TSP_USEC: 993 microtime(&tv); 994 TIMEVAL_TO_TIMESPEC(&tv, tsp); 995 break; 996 case TSP_NSEC: 997 default: 998 nanotime(tsp); 999 break; 1000 } 1001 } 1002 1003 /* 1004 * Set vnode attributes to VNOVAL 1005 */ 1006 void 1007 vattr_null(struct vattr *vap) 1008 { 1009 1010 vap->va_type = VNON; 1011 vap->va_size = VNOVAL; 1012 vap->va_bytes = VNOVAL; 1013 vap->va_mode = VNOVAL; 1014 vap->va_nlink = VNOVAL; 1015 vap->va_uid = VNOVAL; 1016 vap->va_gid = VNOVAL; 1017 vap->va_fsid = VNOVAL; 1018 vap->va_fileid = VNOVAL; 1019 vap->va_blocksize = VNOVAL; 1020 vap->va_rdev = VNOVAL; 1021 vap->va_atime.tv_sec = VNOVAL; 1022 vap->va_atime.tv_nsec = VNOVAL; 1023 vap->va_mtime.tv_sec = VNOVAL; 1024 vap->va_mtime.tv_nsec = VNOVAL; 1025 vap->va_ctime.tv_sec = VNOVAL; 1026 vap->va_ctime.tv_nsec = VNOVAL; 1027 vap->va_birthtime.tv_sec = VNOVAL; 1028 vap->va_birthtime.tv_nsec = VNOVAL; 1029 vap->va_flags = VNOVAL; 1030 vap->va_gen = VNOVAL; 1031 vap->va_vaflags = 0; 1032 } 1033 1034 /* 1035 * This routine is called when we have too many vnodes. It attempts 1036 * to free <count> vnodes and will potentially free vnodes that still 1037 * have VM backing store (VM backing store is typically the cause 1038 * of a vnode blowout so we want to do this). Therefore, this operation 1039 * is not considered cheap. 1040 * 1041 * A number of conditions may prevent a vnode from being reclaimed. 1042 * the buffer cache may have references on the vnode, a directory 1043 * vnode may still have references due to the namei cache representing 1044 * underlying files, or the vnode may be in active use. It is not 1045 * desirable to reuse such vnodes. These conditions may cause the 1046 * number of vnodes to reach some minimum value regardless of what 1047 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 1048 * 1049 * @param mp Try to reclaim vnodes from this mountpoint 1050 * @param reclaim_nc_src Only reclaim directories with outgoing namecache 1051 * entries if this argument is strue 1052 * @param trigger Only reclaim vnodes with fewer than this many resident 1053 * pages. 1054 * @return The number of vnodes that were reclaimed. 1055 */ 1056 static int 1057 vlrureclaim(struct mount *mp, bool reclaim_nc_src, int trigger) 1058 { 1059 struct vnode *vp; 1060 int count, done, target; 1061 1062 done = 0; 1063 vn_start_write(NULL, &mp, V_WAIT); 1064 MNT_ILOCK(mp); 1065 count = mp->mnt_nvnodelistsize; 1066 target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1); 1067 target = target / 10 + 1; 1068 while (count != 0 && done < target) { 1069 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 1070 while (vp != NULL && vp->v_type == VMARKER) 1071 vp = TAILQ_NEXT(vp, v_nmntvnodes); 1072 if (vp == NULL) 1073 break; 1074 /* 1075 * XXX LRU is completely broken for non-free vnodes. First 1076 * by calling here in mountpoint order, then by moving 1077 * unselected vnodes to the end here, and most grossly by 1078 * removing the vlruvp() function that was supposed to 1079 * maintain the order. (This function was born broken 1080 * since syncer problems prevented it doing anything.) The 1081 * order is closer to LRC (C = Created). 1082 * 1083 * LRU reclaiming of vnodes seems to have last worked in 1084 * FreeBSD-3 where LRU wasn't mentioned under any spelling. 1085 * Then there was no hold count, and inactive vnodes were 1086 * simply put on the free list in LRU order. The separate 1087 * lists also break LRU. We prefer to reclaim from the 1088 * free list for technical reasons. This tends to thrash 1089 * the free list to keep very unrecently used held vnodes. 1090 * The problem is mitigated by keeping the free list large. 1091 */ 1092 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1093 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1094 --count; 1095 if (!VI_TRYLOCK(vp)) 1096 goto next_iter; 1097 /* 1098 * If it's been deconstructed already, it's still 1099 * referenced, or it exceeds the trigger, skip it. 1100 * Also skip free vnodes. We are trying to make space 1101 * to expand the free list, not reduce it. 1102 */ 1103 if (vp->v_usecount || 1104 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1105 ((vp->v_iflag & VI_FREE) != 0) || 1106 VN_IS_DOOMED(vp) || (vp->v_object != NULL && 1107 vp->v_object->resident_page_count > trigger)) { 1108 VI_UNLOCK(vp); 1109 goto next_iter; 1110 } 1111 MNT_IUNLOCK(mp); 1112 vholdl(vp); 1113 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) { 1114 vdrop(vp); 1115 goto next_iter_mntunlocked; 1116 } 1117 VI_LOCK(vp); 1118 /* 1119 * v_usecount may have been bumped after VOP_LOCK() dropped 1120 * the vnode interlock and before it was locked again. 1121 * 1122 * It is not necessary to recheck VIRF_DOOMED because it can 1123 * only be set by another thread that holds both the vnode 1124 * lock and vnode interlock. If another thread has the 1125 * vnode lock before we get to VOP_LOCK() and obtains the 1126 * vnode interlock after VOP_LOCK() drops the vnode 1127 * interlock, the other thread will be unable to drop the 1128 * vnode lock before our VOP_LOCK() call fails. 1129 */ 1130 if (vp->v_usecount || 1131 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1132 (vp->v_object != NULL && 1133 vp->v_object->resident_page_count > trigger)) { 1134 VOP_UNLOCK(vp); 1135 vdropl(vp); 1136 goto next_iter_mntunlocked; 1137 } 1138 KASSERT(!VN_IS_DOOMED(vp), 1139 ("VIRF_DOOMED unexpectedly detected in vlrureclaim()")); 1140 counter_u64_add(recycles_count, 1); 1141 vgonel(vp); 1142 VOP_UNLOCK(vp); 1143 vdropl(vp); 1144 done++; 1145 next_iter_mntunlocked: 1146 if (!should_yield()) 1147 goto relock_mnt; 1148 goto yield; 1149 next_iter: 1150 if (!should_yield()) 1151 continue; 1152 MNT_IUNLOCK(mp); 1153 yield: 1154 kern_yield(PRI_USER); 1155 relock_mnt: 1156 MNT_ILOCK(mp); 1157 } 1158 MNT_IUNLOCK(mp); 1159 vn_finished_write(mp); 1160 return done; 1161 } 1162 1163 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */ 1164 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free, 1165 0, 1166 "limit on vnode free requests per call to the vnlru_free routine"); 1167 1168 /* 1169 * Attempt to reduce the free list by the requested amount. 1170 */ 1171 static void 1172 vnlru_free_locked(int count, struct vfsops *mnt_op) 1173 { 1174 struct vnode *vp; 1175 struct mount *mp; 1176 bool tried_batches; 1177 1178 tried_batches = false; 1179 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 1180 if (count > max_vnlru_free) 1181 count = max_vnlru_free; 1182 for (; count > 0; count--) { 1183 vp = TAILQ_FIRST(&vnode_free_list); 1184 /* 1185 * The list can be modified while the free_list_mtx 1186 * has been dropped and vp could be NULL here. 1187 */ 1188 if (vp == NULL) { 1189 if (tried_batches) 1190 break; 1191 mtx_unlock(&vnode_free_list_mtx); 1192 vnlru_return_batches(mnt_op); 1193 tried_batches = true; 1194 mtx_lock(&vnode_free_list_mtx); 1195 continue; 1196 } 1197 1198 VNASSERT(vp->v_op != NULL, vp, 1199 ("vnlru_free: vnode already reclaimed.")); 1200 KASSERT((vp->v_iflag & VI_FREE) != 0, 1201 ("Removing vnode not on freelist")); 1202 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 1203 ("Mangling active vnode")); 1204 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 1205 1206 /* 1207 * Don't recycle if our vnode is from different type 1208 * of mount point. Note that mp is type-safe, the 1209 * check does not reach unmapped address even if 1210 * vnode is reclaimed. 1211 * Don't recycle if we can't get the interlock without 1212 * blocking. 1213 */ 1214 if ((mnt_op != NULL && (mp = vp->v_mount) != NULL && 1215 mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) { 1216 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist); 1217 continue; 1218 } 1219 VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0, 1220 vp, ("vp inconsistent on freelist")); 1221 1222 /* 1223 * The clear of VI_FREE prevents activation of the 1224 * vnode. There is no sense in putting the vnode on 1225 * the mount point active list, only to remove it 1226 * later during recycling. Inline the relevant part 1227 * of vholdl(), to avoid triggering assertions or 1228 * activating. 1229 */ 1230 freevnodes--; 1231 vp->v_iflag &= ~VI_FREE; 1232 VNODE_REFCOUNT_FENCE_REL(); 1233 refcount_acquire(&vp->v_holdcnt); 1234 1235 mtx_unlock(&vnode_free_list_mtx); 1236 VI_UNLOCK(vp); 1237 vtryrecycle(vp); 1238 /* 1239 * If the recycled succeeded this vdrop will actually free 1240 * the vnode. If not it will simply place it back on 1241 * the free list. 1242 */ 1243 vdrop(vp); 1244 mtx_lock(&vnode_free_list_mtx); 1245 } 1246 } 1247 1248 void 1249 vnlru_free(int count, struct vfsops *mnt_op) 1250 { 1251 1252 mtx_lock(&vnode_free_list_mtx); 1253 vnlru_free_locked(count, mnt_op); 1254 mtx_unlock(&vnode_free_list_mtx); 1255 } 1256 1257 static void 1258 vnlru_recalc(void) 1259 { 1260 1261 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 1262 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); 1263 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ 1264 vlowat = vhiwat / 2; 1265 } 1266 1267 /* XXX some names and initialization are bad for limits and watermarks. */ 1268 static int 1269 vspace(void) 1270 { 1271 u_long rnumvnodes, rfreevnodes; 1272 int space; 1273 1274 rnumvnodes = atomic_load_long(&numvnodes); 1275 rfreevnodes = atomic_load_long(&freevnodes); 1276 if (rnumvnodes > desiredvnodes) 1277 return (0); 1278 space = desiredvnodes - rnumvnodes; 1279 if (freevnodes > wantfreevnodes) 1280 space += rfreevnodes - wantfreevnodes; 1281 return (space); 1282 } 1283 1284 static void 1285 vnlru_return_batch_locked(struct mount *mp) 1286 { 1287 struct vnode *vp; 1288 1289 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 1290 1291 if (mp->mnt_tmpfreevnodelistsize == 0) 1292 return; 1293 1294 TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) { 1295 VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp, 1296 ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist")); 1297 vp->v_mflag &= ~VMP_TMPMNTFREELIST; 1298 } 1299 mtx_lock(&vnode_free_list_mtx); 1300 TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist); 1301 freevnodes += mp->mnt_tmpfreevnodelistsize; 1302 mtx_unlock(&vnode_free_list_mtx); 1303 mp->mnt_tmpfreevnodelistsize = 0; 1304 } 1305 1306 static void 1307 vnlru_return_batch(struct mount *mp) 1308 { 1309 1310 mtx_lock(&mp->mnt_listmtx); 1311 vnlru_return_batch_locked(mp); 1312 mtx_unlock(&mp->mnt_listmtx); 1313 } 1314 1315 static void 1316 vnlru_return_batches(struct vfsops *mnt_op) 1317 { 1318 struct mount *mp, *nmp; 1319 bool need_unbusy; 1320 1321 mtx_lock(&mountlist_mtx); 1322 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 1323 need_unbusy = false; 1324 if (mnt_op != NULL && mp->mnt_op != mnt_op) 1325 goto next; 1326 if (mp->mnt_tmpfreevnodelistsize == 0) 1327 goto next; 1328 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) { 1329 vnlru_return_batch(mp); 1330 need_unbusy = true; 1331 mtx_lock(&mountlist_mtx); 1332 } 1333 next: 1334 nmp = TAILQ_NEXT(mp, mnt_list); 1335 if (need_unbusy) 1336 vfs_unbusy(mp); 1337 } 1338 mtx_unlock(&mountlist_mtx); 1339 } 1340 1341 /* 1342 * Attempt to recycle vnodes in a context that is always safe to block. 1343 * Calling vlrurecycle() from the bowels of filesystem code has some 1344 * interesting deadlock problems. 1345 */ 1346 static struct proc *vnlruproc; 1347 static int vnlruproc_sig; 1348 1349 static void 1350 vnlru_proc(void) 1351 { 1352 u_long rnumvnodes, rfreevnodes; 1353 struct mount *mp, *nmp; 1354 unsigned long onumvnodes; 1355 int done, force, trigger, usevnodes, vsp; 1356 bool reclaim_nc_src; 1357 1358 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, 1359 SHUTDOWN_PRI_FIRST); 1360 1361 force = 0; 1362 for (;;) { 1363 kproc_suspend_check(vnlruproc); 1364 mtx_lock(&vnode_free_list_mtx); 1365 rnumvnodes = atomic_load_long(&numvnodes); 1366 /* 1367 * If numvnodes is too large (due to desiredvnodes being 1368 * adjusted using its sysctl, or emergency growth), first 1369 * try to reduce it by discarding from the free list. 1370 */ 1371 if (rnumvnodes > desiredvnodes) 1372 vnlru_free_locked(rnumvnodes - desiredvnodes, NULL); 1373 /* 1374 * Sleep if the vnode cache is in a good state. This is 1375 * when it is not over-full and has space for about a 4% 1376 * or 9% expansion (by growing its size or inexcessively 1377 * reducing its free list). Otherwise, try to reclaim 1378 * space for a 10% expansion. 1379 */ 1380 if (vstir && force == 0) { 1381 force = 1; 1382 vstir = 0; 1383 } 1384 vsp = vspace(); 1385 if (vsp >= vlowat && force == 0) { 1386 vnlruproc_sig = 0; 1387 wakeup(&vnlruproc_sig); 1388 msleep(vnlruproc, &vnode_free_list_mtx, 1389 PVFS|PDROP, "vlruwt", hz); 1390 continue; 1391 } 1392 mtx_unlock(&vnode_free_list_mtx); 1393 done = 0; 1394 rnumvnodes = atomic_load_long(&numvnodes); 1395 rfreevnodes = atomic_load_long(&freevnodes); 1396 1397 onumvnodes = rnumvnodes; 1398 /* 1399 * Calculate parameters for recycling. These are the same 1400 * throughout the loop to give some semblance of fairness. 1401 * The trigger point is to avoid recycling vnodes with lots 1402 * of resident pages. We aren't trying to free memory; we 1403 * are trying to recycle or at least free vnodes. 1404 */ 1405 if (rnumvnodes <= desiredvnodes) 1406 usevnodes = rnumvnodes - rfreevnodes; 1407 else 1408 usevnodes = rnumvnodes; 1409 if (usevnodes <= 0) 1410 usevnodes = 1; 1411 /* 1412 * The trigger value is is chosen to give a conservatively 1413 * large value to ensure that it alone doesn't prevent 1414 * making progress. The value can easily be so large that 1415 * it is effectively infinite in some congested and 1416 * misconfigured cases, and this is necessary. Normally 1417 * it is about 8 to 100 (pages), which is quite large. 1418 */ 1419 trigger = vm_cnt.v_page_count * 2 / usevnodes; 1420 if (force < 2) 1421 trigger = vsmalltrigger; 1422 reclaim_nc_src = force >= 3; 1423 mtx_lock(&mountlist_mtx); 1424 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 1425 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) { 1426 nmp = TAILQ_NEXT(mp, mnt_list); 1427 continue; 1428 } 1429 done += vlrureclaim(mp, reclaim_nc_src, trigger); 1430 mtx_lock(&mountlist_mtx); 1431 nmp = TAILQ_NEXT(mp, mnt_list); 1432 vfs_unbusy(mp); 1433 } 1434 mtx_unlock(&mountlist_mtx); 1435 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes) 1436 uma_reclaim(UMA_RECLAIM_DRAIN); 1437 if (done == 0) { 1438 if (force == 0 || force == 1) { 1439 force = 2; 1440 continue; 1441 } 1442 if (force == 2) { 1443 force = 3; 1444 continue; 1445 } 1446 force = 0; 1447 vnlru_nowhere++; 1448 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 1449 } else 1450 kern_yield(PRI_USER); 1451 /* 1452 * After becoming active to expand above low water, keep 1453 * active until above high water. 1454 */ 1455 vsp = vspace(); 1456 force = vsp < vhiwat; 1457 } 1458 } 1459 1460 static struct kproc_desc vnlru_kp = { 1461 "vnlru", 1462 vnlru_proc, 1463 &vnlruproc 1464 }; 1465 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 1466 &vnlru_kp); 1467 1468 /* 1469 * Routines having to do with the management of the vnode table. 1470 */ 1471 1472 /* 1473 * Try to recycle a freed vnode. We abort if anyone picks up a reference 1474 * before we actually vgone(). This function must be called with the vnode 1475 * held to prevent the vnode from being returned to the free list midway 1476 * through vgone(). 1477 */ 1478 static int 1479 vtryrecycle(struct vnode *vp) 1480 { 1481 struct mount *vnmp; 1482 1483 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 1484 VNASSERT(vp->v_holdcnt, vp, 1485 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 1486 /* 1487 * This vnode may found and locked via some other list, if so we 1488 * can't recycle it yet. 1489 */ 1490 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1491 CTR2(KTR_VFS, 1492 "%s: impossible to recycle, vp %p lock is already held", 1493 __func__, vp); 1494 return (EWOULDBLOCK); 1495 } 1496 /* 1497 * Don't recycle if its filesystem is being suspended. 1498 */ 1499 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 1500 VOP_UNLOCK(vp); 1501 CTR2(KTR_VFS, 1502 "%s: impossible to recycle, cannot start the write for %p", 1503 __func__, vp); 1504 return (EBUSY); 1505 } 1506 /* 1507 * If we got this far, we need to acquire the interlock and see if 1508 * anyone picked up this vnode from another list. If not, we will 1509 * mark it with DOOMED via vgonel() so that anyone who does find it 1510 * will skip over it. 1511 */ 1512 VI_LOCK(vp); 1513 if (vp->v_usecount) { 1514 VOP_UNLOCK(vp); 1515 VI_UNLOCK(vp); 1516 vn_finished_write(vnmp); 1517 CTR2(KTR_VFS, 1518 "%s: impossible to recycle, %p is already referenced", 1519 __func__, vp); 1520 return (EBUSY); 1521 } 1522 if (!VN_IS_DOOMED(vp)) { 1523 counter_u64_add(recycles_free_count, 1); 1524 vgonel(vp); 1525 } 1526 VOP_UNLOCK(vp); 1527 VI_UNLOCK(vp); 1528 vn_finished_write(vnmp); 1529 return (0); 1530 } 1531 1532 static void 1533 vcheckspace(void) 1534 { 1535 int vsp; 1536 1537 vsp = vspace(); 1538 if (vsp < vlowat && vnlruproc_sig == 0) { 1539 vnlruproc_sig = 1; 1540 wakeup(vnlruproc); 1541 } 1542 } 1543 1544 /* 1545 * Wait if necessary for space for a new vnode. 1546 */ 1547 static int 1548 vn_alloc_wait(int suspended) 1549 { 1550 1551 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 1552 if (numvnodes >= desiredvnodes) { 1553 if (suspended) { 1554 /* 1555 * The file system is being suspended. We cannot 1556 * risk a deadlock here, so allow allocation of 1557 * another vnode even if this would give too many. 1558 */ 1559 return (0); 1560 } 1561 if (vnlruproc_sig == 0) { 1562 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 1563 wakeup(vnlruproc); 1564 } 1565 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, 1566 "vlruwk", hz); 1567 } 1568 /* Post-adjust like the pre-adjust in getnewvnode(). */ 1569 if (numvnodes + 1 > desiredvnodes && freevnodes > 1) 1570 vnlru_free_locked(1, NULL); 1571 return (numvnodes >= desiredvnodes ? ENFILE : 0); 1572 } 1573 1574 static struct vnode * 1575 vn_alloc(struct mount *mp) 1576 { 1577 static int cyclecount; 1578 int error __unused; 1579 1580 mtx_lock(&vnode_free_list_mtx); 1581 if (numvnodes < desiredvnodes) 1582 cyclecount = 0; 1583 else if (cyclecount++ >= freevnodes) { 1584 cyclecount = 0; 1585 vstir = 1; 1586 } 1587 /* 1588 * Grow the vnode cache if it will not be above its target max 1589 * after growing. Otherwise, if the free list is nonempty, try 1590 * to reclaim 1 item from it before growing the cache (possibly 1591 * above its target max if the reclamation failed or is delayed). 1592 * Otherwise, wait for some space. In all cases, schedule 1593 * vnlru_proc() if we are getting short of space. The watermarks 1594 * should be chosen so that we never wait or even reclaim from 1595 * the free list to below its target minimum. 1596 */ 1597 if (numvnodes + 1 <= desiredvnodes) 1598 ; 1599 else if (freevnodes > 0) 1600 vnlru_free_locked(1, NULL); 1601 else { 1602 error = vn_alloc_wait(mp != NULL && (mp->mnt_kern_flag & 1603 MNTK_SUSPEND)); 1604 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */ 1605 if (error != 0) { 1606 mtx_unlock(&vnode_free_list_mtx); 1607 return (error); 1608 } 1609 #endif 1610 } 1611 vcheckspace(); 1612 atomic_add_long(&numvnodes, 1); 1613 mtx_unlock(&vnode_free_list_mtx); 1614 return (uma_zalloc(vnode_zone, M_WAITOK)); 1615 } 1616 1617 static void 1618 vn_free(struct vnode *vp) 1619 { 1620 1621 atomic_subtract_long(&numvnodes, 1); 1622 uma_zfree(vnode_zone, vp); 1623 } 1624 1625 /* 1626 * Return the next vnode from the free list. 1627 */ 1628 int 1629 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 1630 struct vnode **vpp) 1631 { 1632 struct vnode *vp; 1633 struct thread *td; 1634 struct lock_object *lo; 1635 1636 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 1637 1638 KASSERT(vops->registered, 1639 ("%s: not registered vector op %p\n", __func__, vops)); 1640 1641 td = curthread; 1642 if (td->td_vp_reserved != NULL) { 1643 vp = td->td_vp_reserved; 1644 td->td_vp_reserved = NULL; 1645 } else { 1646 vp = vn_alloc(mp); 1647 } 1648 counter_u64_add(vnodes_created, 1); 1649 /* 1650 * Locks are given the generic name "vnode" when created. 1651 * Follow the historic practice of using the filesystem 1652 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. 1653 * 1654 * Locks live in a witness group keyed on their name. Thus, 1655 * when a lock is renamed, it must also move from the witness 1656 * group of its old name to the witness group of its new name. 1657 * 1658 * The change only needs to be made when the vnode moves 1659 * from one filesystem type to another. We ensure that each 1660 * filesystem use a single static name pointer for its tag so 1661 * that we can compare pointers rather than doing a strcmp(). 1662 */ 1663 lo = &vp->v_vnlock->lock_object; 1664 if (lo->lo_name != tag) { 1665 lo->lo_name = tag; 1666 WITNESS_DESTROY(lo); 1667 WITNESS_INIT(lo, tag); 1668 } 1669 /* 1670 * By default, don't allow shared locks unless filesystems opt-in. 1671 */ 1672 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; 1673 /* 1674 * Finalize various vnode identity bits. 1675 */ 1676 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); 1677 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); 1678 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); 1679 vp->v_type = VNON; 1680 vp->v_op = vops; 1681 v_init_counters(vp); 1682 vp->v_bufobj.bo_ops = &buf_ops_bio; 1683 #ifdef DIAGNOSTIC 1684 if (mp == NULL && vops != &dead_vnodeops) 1685 printf("NULL mp in getnewvnode(9), tag %s\n", tag); 1686 #endif 1687 #ifdef MAC 1688 mac_vnode_init(vp); 1689 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1690 mac_vnode_associate_singlelabel(mp, vp); 1691 #endif 1692 if (mp != NULL) { 1693 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 1694 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 1695 vp->v_vflag |= VV_NOKNOTE; 1696 } 1697 1698 /* 1699 * For the filesystems which do not use vfs_hash_insert(), 1700 * still initialize v_hash to have vfs_hash_index() useful. 1701 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 1702 * its own hashing. 1703 */ 1704 vp->v_hash = (uintptr_t)vp >> vnsz2log; 1705 1706 *vpp = vp; 1707 return (0); 1708 } 1709 1710 void 1711 getnewvnode_reserve(void) 1712 { 1713 struct thread *td; 1714 1715 td = curthread; 1716 MPASS(td->td_vp_reserved == NULL); 1717 td->td_vp_reserved = vn_alloc(NULL); 1718 } 1719 1720 void 1721 getnewvnode_drop_reserve(void) 1722 { 1723 struct thread *td; 1724 1725 td = curthread; 1726 if (td->td_vp_reserved != NULL) { 1727 vn_free(td->td_vp_reserved); 1728 td->td_vp_reserved = NULL; 1729 } 1730 } 1731 1732 static void 1733 freevnode(struct vnode *vp) 1734 { 1735 struct bufobj *bo; 1736 1737 /* 1738 * The vnode has been marked for destruction, so free it. 1739 * 1740 * The vnode will be returned to the zone where it will 1741 * normally remain until it is needed for another vnode. We 1742 * need to cleanup (or verify that the cleanup has already 1743 * been done) any residual data left from its current use 1744 * so as not to contaminate the freshly allocated vnode. 1745 */ 1746 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 1747 bo = &vp->v_bufobj; 1748 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 1749 ("cleaned vnode still on the free list.")); 1750 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 1751 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 1752 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 1753 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 1754 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 1755 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 1756 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 1757 ("clean blk trie not empty")); 1758 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 1759 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 1760 ("dirty blk trie not empty")); 1761 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 1762 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 1763 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); 1764 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, 1765 ("Dangling rangelock waiters")); 1766 VI_UNLOCK(vp); 1767 #ifdef MAC 1768 mac_vnode_destroy(vp); 1769 #endif 1770 if (vp->v_pollinfo != NULL) { 1771 destroy_vpollinfo(vp->v_pollinfo); 1772 vp->v_pollinfo = NULL; 1773 } 1774 #ifdef INVARIANTS 1775 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ 1776 vp->v_op = NULL; 1777 #endif 1778 vp->v_mountedhere = NULL; 1779 vp->v_unpcb = NULL; 1780 vp->v_rdev = NULL; 1781 vp->v_fifoinfo = NULL; 1782 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 1783 vp->v_irflag = 0; 1784 vp->v_iflag = 0; 1785 vp->v_vflag = 0; 1786 bo->bo_flag = 0; 1787 vn_free(vp); 1788 } 1789 1790 /* 1791 * Delete from old mount point vnode list, if on one. 1792 */ 1793 static void 1794 delmntque(struct vnode *vp) 1795 { 1796 struct mount *mp; 1797 1798 mp = vp->v_mount; 1799 if (mp == NULL) 1800 return; 1801 MNT_ILOCK(mp); 1802 VI_LOCK(vp); 1803 KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize, 1804 ("Active vnode list size %d > Vnode list size %d", 1805 mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize)); 1806 if (vp->v_iflag & VI_ACTIVE) { 1807 vp->v_iflag &= ~VI_ACTIVE; 1808 mtx_lock(&mp->mnt_listmtx); 1809 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist); 1810 mp->mnt_activevnodelistsize--; 1811 mtx_unlock(&mp->mnt_listmtx); 1812 } 1813 vp->v_mount = NULL; 1814 VI_UNLOCK(vp); 1815 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1816 ("bad mount point vnode list size")); 1817 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1818 mp->mnt_nvnodelistsize--; 1819 MNT_REL(mp); 1820 MNT_IUNLOCK(mp); 1821 } 1822 1823 static void 1824 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 1825 { 1826 1827 vp->v_data = NULL; 1828 vp->v_op = &dead_vnodeops; 1829 vgone(vp); 1830 vput(vp); 1831 } 1832 1833 /* 1834 * Insert into list of vnodes for the new mount point, if available. 1835 */ 1836 int 1837 insmntque1(struct vnode *vp, struct mount *mp, 1838 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1839 { 1840 1841 KASSERT(vp->v_mount == NULL, 1842 ("insmntque: vnode already on per mount vnode list")); 1843 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1844 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 1845 1846 /* 1847 * We acquire the vnode interlock early to ensure that the 1848 * vnode cannot be recycled by another process releasing a 1849 * holdcnt on it before we get it on both the vnode list 1850 * and the active vnode list. The mount mutex protects only 1851 * manipulation of the vnode list and the vnode freelist 1852 * mutex protects only manipulation of the active vnode list. 1853 * Hence the need to hold the vnode interlock throughout. 1854 */ 1855 MNT_ILOCK(mp); 1856 VI_LOCK(vp); 1857 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 && 1858 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 1859 mp->mnt_nvnodelistsize == 0)) && 1860 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 1861 VI_UNLOCK(vp); 1862 MNT_IUNLOCK(mp); 1863 if (dtr != NULL) 1864 dtr(vp, dtr_arg); 1865 return (EBUSY); 1866 } 1867 vp->v_mount = mp; 1868 MNT_REF(mp); 1869 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1870 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1871 ("neg mount point vnode list size")); 1872 mp->mnt_nvnodelistsize++; 1873 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 1874 ("Activating already active vnode")); 1875 vp->v_iflag |= VI_ACTIVE; 1876 mtx_lock(&mp->mnt_listmtx); 1877 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 1878 mp->mnt_activevnodelistsize++; 1879 mtx_unlock(&mp->mnt_listmtx); 1880 VI_UNLOCK(vp); 1881 MNT_IUNLOCK(mp); 1882 return (0); 1883 } 1884 1885 int 1886 insmntque(struct vnode *vp, struct mount *mp) 1887 { 1888 1889 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1890 } 1891 1892 /* 1893 * Flush out and invalidate all buffers associated with a bufobj 1894 * Called with the underlying object locked. 1895 */ 1896 int 1897 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 1898 { 1899 int error; 1900 1901 BO_LOCK(bo); 1902 if (flags & V_SAVE) { 1903 error = bufobj_wwait(bo, slpflag, slptimeo); 1904 if (error) { 1905 BO_UNLOCK(bo); 1906 return (error); 1907 } 1908 if (bo->bo_dirty.bv_cnt > 0) { 1909 BO_UNLOCK(bo); 1910 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) 1911 return (error); 1912 /* 1913 * XXX We could save a lock/unlock if this was only 1914 * enabled under INVARIANTS 1915 */ 1916 BO_LOCK(bo); 1917 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1918 panic("vinvalbuf: dirty bufs"); 1919 } 1920 } 1921 /* 1922 * If you alter this loop please notice that interlock is dropped and 1923 * reacquired in flushbuflist. Special care is needed to ensure that 1924 * no race conditions occur from this. 1925 */ 1926 do { 1927 error = flushbuflist(&bo->bo_clean, 1928 flags, bo, slpflag, slptimeo); 1929 if (error == 0 && !(flags & V_CLEANONLY)) 1930 error = flushbuflist(&bo->bo_dirty, 1931 flags, bo, slpflag, slptimeo); 1932 if (error != 0 && error != EAGAIN) { 1933 BO_UNLOCK(bo); 1934 return (error); 1935 } 1936 } while (error != 0); 1937 1938 /* 1939 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1940 * have write I/O in-progress but if there is a VM object then the 1941 * VM object can also have read-I/O in-progress. 1942 */ 1943 do { 1944 bufobj_wwait(bo, 0, 0); 1945 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) { 1946 BO_UNLOCK(bo); 1947 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx"); 1948 BO_LOCK(bo); 1949 } 1950 } while (bo->bo_numoutput > 0); 1951 BO_UNLOCK(bo); 1952 1953 /* 1954 * Destroy the copy in the VM cache, too. 1955 */ 1956 if (bo->bo_object != NULL && 1957 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { 1958 VM_OBJECT_WLOCK(bo->bo_object); 1959 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 1960 OBJPR_CLEANONLY : 0); 1961 VM_OBJECT_WUNLOCK(bo->bo_object); 1962 } 1963 1964 #ifdef INVARIANTS 1965 BO_LOCK(bo); 1966 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO | 1967 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 || 1968 bo->bo_clean.bv_cnt > 0)) 1969 panic("vinvalbuf: flush failed"); 1970 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && 1971 bo->bo_dirty.bv_cnt > 0) 1972 panic("vinvalbuf: flush dirty failed"); 1973 BO_UNLOCK(bo); 1974 #endif 1975 return (0); 1976 } 1977 1978 /* 1979 * Flush out and invalidate all buffers associated with a vnode. 1980 * Called with the underlying object locked. 1981 */ 1982 int 1983 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 1984 { 1985 1986 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 1987 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1988 if (vp->v_object != NULL && vp->v_object->handle != vp) 1989 return (0); 1990 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 1991 } 1992 1993 /* 1994 * Flush out buffers on the specified list. 1995 * 1996 */ 1997 static int 1998 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1999 int slptimeo) 2000 { 2001 struct buf *bp, *nbp; 2002 int retval, error; 2003 daddr_t lblkno; 2004 b_xflags_t xflags; 2005 2006 ASSERT_BO_WLOCKED(bo); 2007 2008 retval = 0; 2009 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 2010 /* 2011 * If we are flushing both V_NORMAL and V_ALT buffers then 2012 * do not skip any buffers. If we are flushing only V_NORMAL 2013 * buffers then skip buffers marked as BX_ALTDATA. If we are 2014 * flushing only V_ALT buffers then skip buffers not marked 2015 * as BX_ALTDATA. 2016 */ 2017 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) && 2018 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) || 2019 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) { 2020 continue; 2021 } 2022 if (nbp != NULL) { 2023 lblkno = nbp->b_lblkno; 2024 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 2025 } 2026 retval = EAGAIN; 2027 error = BUF_TIMELOCK(bp, 2028 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 2029 "flushbuf", slpflag, slptimeo); 2030 if (error) { 2031 BO_LOCK(bo); 2032 return (error != ENOLCK ? error : EAGAIN); 2033 } 2034 KASSERT(bp->b_bufobj == bo, 2035 ("bp %p wrong b_bufobj %p should be %p", 2036 bp, bp->b_bufobj, bo)); 2037 /* 2038 * XXX Since there are no node locks for NFS, I 2039 * believe there is a slight chance that a delayed 2040 * write will occur while sleeping just above, so 2041 * check for it. 2042 */ 2043 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 2044 (flags & V_SAVE)) { 2045 bremfree(bp); 2046 bp->b_flags |= B_ASYNC; 2047 bwrite(bp); 2048 BO_LOCK(bo); 2049 return (EAGAIN); /* XXX: why not loop ? */ 2050 } 2051 bremfree(bp); 2052 bp->b_flags |= (B_INVAL | B_RELBUF); 2053 bp->b_flags &= ~B_ASYNC; 2054 brelse(bp); 2055 BO_LOCK(bo); 2056 if (nbp == NULL) 2057 break; 2058 nbp = gbincore(bo, lblkno); 2059 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2060 != xflags) 2061 break; /* nbp invalid */ 2062 } 2063 return (retval); 2064 } 2065 2066 int 2067 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 2068 { 2069 struct buf *bp; 2070 int error; 2071 daddr_t lblkno; 2072 2073 ASSERT_BO_LOCKED(bo); 2074 2075 for (lblkno = startn;;) { 2076 again: 2077 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); 2078 if (bp == NULL || bp->b_lblkno >= endn || 2079 bp->b_lblkno < startn) 2080 break; 2081 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 2082 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 2083 if (error != 0) { 2084 BO_RLOCK(bo); 2085 if (error == ENOLCK) 2086 goto again; 2087 return (error); 2088 } 2089 KASSERT(bp->b_bufobj == bo, 2090 ("bp %p wrong b_bufobj %p should be %p", 2091 bp, bp->b_bufobj, bo)); 2092 lblkno = bp->b_lblkno + 1; 2093 if ((bp->b_flags & B_MANAGED) == 0) 2094 bremfree(bp); 2095 bp->b_flags |= B_RELBUF; 2096 /* 2097 * In the VMIO case, use the B_NOREUSE flag to hint that the 2098 * pages backing each buffer in the range are unlikely to be 2099 * reused. Dirty buffers will have the hint applied once 2100 * they've been written. 2101 */ 2102 if ((bp->b_flags & B_VMIO) != 0) 2103 bp->b_flags |= B_NOREUSE; 2104 brelse(bp); 2105 BO_RLOCK(bo); 2106 } 2107 return (0); 2108 } 2109 2110 /* 2111 * Truncate a file's buffer and pages to a specified length. This 2112 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 2113 * sync activity. 2114 */ 2115 int 2116 vtruncbuf(struct vnode *vp, off_t length, int blksize) 2117 { 2118 struct buf *bp, *nbp; 2119 struct bufobj *bo; 2120 daddr_t startlbn; 2121 2122 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__, 2123 vp, blksize, (uintmax_t)length); 2124 2125 /* 2126 * Round up to the *next* lbn. 2127 */ 2128 startlbn = howmany(length, blksize); 2129 2130 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 2131 2132 bo = &vp->v_bufobj; 2133 restart_unlocked: 2134 BO_LOCK(bo); 2135 2136 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN) 2137 ; 2138 2139 if (length > 0) { 2140 restartsync: 2141 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2142 if (bp->b_lblkno > 0) 2143 continue; 2144 /* 2145 * Since we hold the vnode lock this should only 2146 * fail if we're racing with the buf daemon. 2147 */ 2148 if (BUF_LOCK(bp, 2149 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2150 BO_LOCKPTR(bo)) == ENOLCK) 2151 goto restart_unlocked; 2152 2153 VNASSERT((bp->b_flags & B_DELWRI), vp, 2154 ("buf(%p) on dirty queue without DELWRI", bp)); 2155 2156 bremfree(bp); 2157 bawrite(bp); 2158 BO_LOCK(bo); 2159 goto restartsync; 2160 } 2161 } 2162 2163 bufobj_wwait(bo, 0, 0); 2164 BO_UNLOCK(bo); 2165 vnode_pager_setsize(vp, length); 2166 2167 return (0); 2168 } 2169 2170 /* 2171 * Invalidate the cached pages of a file's buffer within the range of block 2172 * numbers [startlbn, endlbn). 2173 */ 2174 void 2175 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, 2176 int blksize) 2177 { 2178 struct bufobj *bo; 2179 off_t start, end; 2180 2181 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range"); 2182 2183 start = blksize * startlbn; 2184 end = blksize * endlbn; 2185 2186 bo = &vp->v_bufobj; 2187 BO_LOCK(bo); 2188 MPASS(blksize == bo->bo_bsize); 2189 2190 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN) 2191 ; 2192 2193 BO_UNLOCK(bo); 2194 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1)); 2195 } 2196 2197 static int 2198 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 2199 daddr_t startlbn, daddr_t endlbn) 2200 { 2201 struct buf *bp, *nbp; 2202 bool anyfreed; 2203 2204 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked"); 2205 ASSERT_BO_LOCKED(bo); 2206 2207 do { 2208 anyfreed = false; 2209 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 2210 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2211 continue; 2212 if (BUF_LOCK(bp, 2213 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2214 BO_LOCKPTR(bo)) == ENOLCK) { 2215 BO_LOCK(bo); 2216 return (EAGAIN); 2217 } 2218 2219 bremfree(bp); 2220 bp->b_flags |= B_INVAL | B_RELBUF; 2221 bp->b_flags &= ~B_ASYNC; 2222 brelse(bp); 2223 anyfreed = true; 2224 2225 BO_LOCK(bo); 2226 if (nbp != NULL && 2227 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 2228 nbp->b_vp != vp || 2229 (nbp->b_flags & B_DELWRI) != 0)) 2230 return (EAGAIN); 2231 } 2232 2233 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2234 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2235 continue; 2236 if (BUF_LOCK(bp, 2237 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2238 BO_LOCKPTR(bo)) == ENOLCK) { 2239 BO_LOCK(bo); 2240 return (EAGAIN); 2241 } 2242 bremfree(bp); 2243 bp->b_flags |= B_INVAL | B_RELBUF; 2244 bp->b_flags &= ~B_ASYNC; 2245 brelse(bp); 2246 anyfreed = true; 2247 2248 BO_LOCK(bo); 2249 if (nbp != NULL && 2250 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 2251 (nbp->b_vp != vp) || 2252 (nbp->b_flags & B_DELWRI) == 0)) 2253 return (EAGAIN); 2254 } 2255 } while (anyfreed); 2256 return (0); 2257 } 2258 2259 static void 2260 buf_vlist_remove(struct buf *bp) 2261 { 2262 struct bufv *bv; 2263 2264 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2265 ASSERT_BO_WLOCKED(bp->b_bufobj); 2266 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 2267 (BX_VNDIRTY|BX_VNCLEAN), 2268 ("buf_vlist_remove: Buf %p is on two lists", bp)); 2269 if (bp->b_xflags & BX_VNDIRTY) 2270 bv = &bp->b_bufobj->bo_dirty; 2271 else 2272 bv = &bp->b_bufobj->bo_clean; 2273 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 2274 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 2275 bv->bv_cnt--; 2276 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 2277 } 2278 2279 /* 2280 * Add the buffer to the sorted clean or dirty block list. 2281 * 2282 * NOTE: xflags is passed as a constant, optimizing this inline function! 2283 */ 2284 static void 2285 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 2286 { 2287 struct bufv *bv; 2288 struct buf *n; 2289 int error; 2290 2291 ASSERT_BO_WLOCKED(bo); 2292 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 2293 ("dead bo %p", bo)); 2294 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 2295 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 2296 bp->b_xflags |= xflags; 2297 if (xflags & BX_VNDIRTY) 2298 bv = &bo->bo_dirty; 2299 else 2300 bv = &bo->bo_clean; 2301 2302 /* 2303 * Keep the list ordered. Optimize empty list insertion. Assume 2304 * we tend to grow at the tail so lookup_le should usually be cheaper 2305 * than _ge. 2306 */ 2307 if (bv->bv_cnt == 0 || 2308 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) 2309 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 2310 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) 2311 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 2312 else 2313 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 2314 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); 2315 if (error) 2316 panic("buf_vlist_add: Preallocated nodes insufficient."); 2317 bv->bv_cnt++; 2318 } 2319 2320 /* 2321 * Look up a buffer using the buffer tries. 2322 */ 2323 struct buf * 2324 gbincore(struct bufobj *bo, daddr_t lblkno) 2325 { 2326 struct buf *bp; 2327 2328 ASSERT_BO_LOCKED(bo); 2329 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 2330 if (bp != NULL) 2331 return (bp); 2332 return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno); 2333 } 2334 2335 /* 2336 * Associate a buffer with a vnode. 2337 */ 2338 void 2339 bgetvp(struct vnode *vp, struct buf *bp) 2340 { 2341 struct bufobj *bo; 2342 2343 bo = &vp->v_bufobj; 2344 ASSERT_BO_WLOCKED(bo); 2345 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 2346 2347 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 2348 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 2349 ("bgetvp: bp already attached! %p", bp)); 2350 2351 vhold(vp); 2352 bp->b_vp = vp; 2353 bp->b_bufobj = bo; 2354 /* 2355 * Insert onto list for new vnode. 2356 */ 2357 buf_vlist_add(bp, bo, BX_VNCLEAN); 2358 } 2359 2360 /* 2361 * Disassociate a buffer from a vnode. 2362 */ 2363 void 2364 brelvp(struct buf *bp) 2365 { 2366 struct bufobj *bo; 2367 struct vnode *vp; 2368 2369 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2370 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 2371 2372 /* 2373 * Delete from old vnode list, if on one. 2374 */ 2375 vp = bp->b_vp; /* XXX */ 2376 bo = bp->b_bufobj; 2377 BO_LOCK(bo); 2378 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2379 buf_vlist_remove(bp); 2380 else 2381 panic("brelvp: Buffer %p not on queue.", bp); 2382 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2383 bo->bo_flag &= ~BO_ONWORKLST; 2384 mtx_lock(&sync_mtx); 2385 LIST_REMOVE(bo, bo_synclist); 2386 syncer_worklist_len--; 2387 mtx_unlock(&sync_mtx); 2388 } 2389 bp->b_vp = NULL; 2390 bp->b_bufobj = NULL; 2391 BO_UNLOCK(bo); 2392 vdrop(vp); 2393 } 2394 2395 /* 2396 * Add an item to the syncer work queue. 2397 */ 2398 static void 2399 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 2400 { 2401 int slot; 2402 2403 ASSERT_BO_WLOCKED(bo); 2404 2405 mtx_lock(&sync_mtx); 2406 if (bo->bo_flag & BO_ONWORKLST) 2407 LIST_REMOVE(bo, bo_synclist); 2408 else { 2409 bo->bo_flag |= BO_ONWORKLST; 2410 syncer_worklist_len++; 2411 } 2412 2413 if (delay > syncer_maxdelay - 2) 2414 delay = syncer_maxdelay - 2; 2415 slot = (syncer_delayno + delay) & syncer_mask; 2416 2417 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 2418 mtx_unlock(&sync_mtx); 2419 } 2420 2421 static int 2422 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 2423 { 2424 int error, len; 2425 2426 mtx_lock(&sync_mtx); 2427 len = syncer_worklist_len - sync_vnode_count; 2428 mtx_unlock(&sync_mtx); 2429 error = SYSCTL_OUT(req, &len, sizeof(len)); 2430 return (error); 2431 } 2432 2433 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, 2434 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0, 2435 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 2436 2437 static struct proc *updateproc; 2438 static void sched_sync(void); 2439 static struct kproc_desc up_kp = { 2440 "syncer", 2441 sched_sync, 2442 &updateproc 2443 }; 2444 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 2445 2446 static int 2447 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2448 { 2449 struct vnode *vp; 2450 struct mount *mp; 2451 2452 *bo = LIST_FIRST(slp); 2453 if (*bo == NULL) 2454 return (0); 2455 vp = bo2vnode(*bo); 2456 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2457 return (1); 2458 /* 2459 * We use vhold in case the vnode does not 2460 * successfully sync. vhold prevents the vnode from 2461 * going away when we unlock the sync_mtx so that 2462 * we can acquire the vnode interlock. 2463 */ 2464 vholdl(vp); 2465 mtx_unlock(&sync_mtx); 2466 VI_UNLOCK(vp); 2467 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2468 vdrop(vp); 2469 mtx_lock(&sync_mtx); 2470 return (*bo == LIST_FIRST(slp)); 2471 } 2472 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2473 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2474 VOP_UNLOCK(vp); 2475 vn_finished_write(mp); 2476 BO_LOCK(*bo); 2477 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2478 /* 2479 * Put us back on the worklist. The worklist 2480 * routine will remove us from our current 2481 * position and then add us back in at a later 2482 * position. 2483 */ 2484 vn_syncer_add_to_worklist(*bo, syncdelay); 2485 } 2486 BO_UNLOCK(*bo); 2487 vdrop(vp); 2488 mtx_lock(&sync_mtx); 2489 return (0); 2490 } 2491 2492 static int first_printf = 1; 2493 2494 /* 2495 * System filesystem synchronizer daemon. 2496 */ 2497 static void 2498 sched_sync(void) 2499 { 2500 struct synclist *next, *slp; 2501 struct bufobj *bo; 2502 long starttime; 2503 struct thread *td = curthread; 2504 int last_work_seen; 2505 int net_worklist_len; 2506 int syncer_final_iter; 2507 int error; 2508 2509 last_work_seen = 0; 2510 syncer_final_iter = 0; 2511 syncer_state = SYNCER_RUNNING; 2512 starttime = time_uptime; 2513 td->td_pflags |= TDP_NORUNNINGBUF; 2514 2515 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 2516 SHUTDOWN_PRI_LAST); 2517 2518 mtx_lock(&sync_mtx); 2519 for (;;) { 2520 if (syncer_state == SYNCER_FINAL_DELAY && 2521 syncer_final_iter == 0) { 2522 mtx_unlock(&sync_mtx); 2523 kproc_suspend_check(td->td_proc); 2524 mtx_lock(&sync_mtx); 2525 } 2526 net_worklist_len = syncer_worklist_len - sync_vnode_count; 2527 if (syncer_state != SYNCER_RUNNING && 2528 starttime != time_uptime) { 2529 if (first_printf) { 2530 printf("\nSyncing disks, vnodes remaining... "); 2531 first_printf = 0; 2532 } 2533 printf("%d ", net_worklist_len); 2534 } 2535 starttime = time_uptime; 2536 2537 /* 2538 * Push files whose dirty time has expired. Be careful 2539 * of interrupt race on slp queue. 2540 * 2541 * Skip over empty worklist slots when shutting down. 2542 */ 2543 do { 2544 slp = &syncer_workitem_pending[syncer_delayno]; 2545 syncer_delayno += 1; 2546 if (syncer_delayno == syncer_maxdelay) 2547 syncer_delayno = 0; 2548 next = &syncer_workitem_pending[syncer_delayno]; 2549 /* 2550 * If the worklist has wrapped since the 2551 * it was emptied of all but syncer vnodes, 2552 * switch to the FINAL_DELAY state and run 2553 * for one more second. 2554 */ 2555 if (syncer_state == SYNCER_SHUTTING_DOWN && 2556 net_worklist_len == 0 && 2557 last_work_seen == syncer_delayno) { 2558 syncer_state = SYNCER_FINAL_DELAY; 2559 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 2560 } 2561 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 2562 syncer_worklist_len > 0); 2563 2564 /* 2565 * Keep track of the last time there was anything 2566 * on the worklist other than syncer vnodes. 2567 * Return to the SHUTTING_DOWN state if any 2568 * new work appears. 2569 */ 2570 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 2571 last_work_seen = syncer_delayno; 2572 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 2573 syncer_state = SYNCER_SHUTTING_DOWN; 2574 while (!LIST_EMPTY(slp)) { 2575 error = sync_vnode(slp, &bo, td); 2576 if (error == 1) { 2577 LIST_REMOVE(bo, bo_synclist); 2578 LIST_INSERT_HEAD(next, bo, bo_synclist); 2579 continue; 2580 } 2581 2582 if (first_printf == 0) { 2583 /* 2584 * Drop the sync mutex, because some watchdog 2585 * drivers need to sleep while patting 2586 */ 2587 mtx_unlock(&sync_mtx); 2588 wdog_kern_pat(WD_LASTVAL); 2589 mtx_lock(&sync_mtx); 2590 } 2591 2592 } 2593 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 2594 syncer_final_iter--; 2595 /* 2596 * The variable rushjob allows the kernel to speed up the 2597 * processing of the filesystem syncer process. A rushjob 2598 * value of N tells the filesystem syncer to process the next 2599 * N seconds worth of work on its queue ASAP. Currently rushjob 2600 * is used by the soft update code to speed up the filesystem 2601 * syncer process when the incore state is getting so far 2602 * ahead of the disk that the kernel memory pool is being 2603 * threatened with exhaustion. 2604 */ 2605 if (rushjob > 0) { 2606 rushjob -= 1; 2607 continue; 2608 } 2609 /* 2610 * Just sleep for a short period of time between 2611 * iterations when shutting down to allow some I/O 2612 * to happen. 2613 * 2614 * If it has taken us less than a second to process the 2615 * current work, then wait. Otherwise start right over 2616 * again. We can still lose time if any single round 2617 * takes more than two seconds, but it does not really 2618 * matter as we are just trying to generally pace the 2619 * filesystem activity. 2620 */ 2621 if (syncer_state != SYNCER_RUNNING || 2622 time_uptime == starttime) { 2623 thread_lock(td); 2624 sched_prio(td, PPAUSE); 2625 thread_unlock(td); 2626 } 2627 if (syncer_state != SYNCER_RUNNING) 2628 cv_timedwait(&sync_wakeup, &sync_mtx, 2629 hz / SYNCER_SHUTDOWN_SPEEDUP); 2630 else if (time_uptime == starttime) 2631 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 2632 } 2633 } 2634 2635 /* 2636 * Request the syncer daemon to speed up its work. 2637 * We never push it to speed up more than half of its 2638 * normal turn time, otherwise it could take over the cpu. 2639 */ 2640 int 2641 speedup_syncer(void) 2642 { 2643 int ret = 0; 2644 2645 mtx_lock(&sync_mtx); 2646 if (rushjob < syncdelay / 2) { 2647 rushjob += 1; 2648 stat_rush_requests += 1; 2649 ret = 1; 2650 } 2651 mtx_unlock(&sync_mtx); 2652 cv_broadcast(&sync_wakeup); 2653 return (ret); 2654 } 2655 2656 /* 2657 * Tell the syncer to speed up its work and run though its work 2658 * list several times, then tell it to shut down. 2659 */ 2660 static void 2661 syncer_shutdown(void *arg, int howto) 2662 { 2663 2664 if (howto & RB_NOSYNC) 2665 return; 2666 mtx_lock(&sync_mtx); 2667 syncer_state = SYNCER_SHUTTING_DOWN; 2668 rushjob = 0; 2669 mtx_unlock(&sync_mtx); 2670 cv_broadcast(&sync_wakeup); 2671 kproc_shutdown(arg, howto); 2672 } 2673 2674 void 2675 syncer_suspend(void) 2676 { 2677 2678 syncer_shutdown(updateproc, 0); 2679 } 2680 2681 void 2682 syncer_resume(void) 2683 { 2684 2685 mtx_lock(&sync_mtx); 2686 first_printf = 1; 2687 syncer_state = SYNCER_RUNNING; 2688 mtx_unlock(&sync_mtx); 2689 cv_broadcast(&sync_wakeup); 2690 kproc_resume(updateproc); 2691 } 2692 2693 /* 2694 * Reassign a buffer from one vnode to another. 2695 * Used to assign file specific control information 2696 * (indirect blocks) to the vnode to which they belong. 2697 */ 2698 void 2699 reassignbuf(struct buf *bp) 2700 { 2701 struct vnode *vp; 2702 struct bufobj *bo; 2703 int delay; 2704 #ifdef INVARIANTS 2705 struct bufv *bv; 2706 #endif 2707 2708 vp = bp->b_vp; 2709 bo = bp->b_bufobj; 2710 ++reassignbufcalls; 2711 2712 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 2713 bp, bp->b_vp, bp->b_flags); 2714 /* 2715 * B_PAGING flagged buffers cannot be reassigned because their vp 2716 * is not fully linked in. 2717 */ 2718 if (bp->b_flags & B_PAGING) 2719 panic("cannot reassign paging buffer"); 2720 2721 /* 2722 * Delete from old vnode list, if on one. 2723 */ 2724 BO_LOCK(bo); 2725 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2726 buf_vlist_remove(bp); 2727 else 2728 panic("reassignbuf: Buffer %p not on queue.", bp); 2729 /* 2730 * If dirty, put on list of dirty buffers; otherwise insert onto list 2731 * of clean buffers. 2732 */ 2733 if (bp->b_flags & B_DELWRI) { 2734 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 2735 switch (vp->v_type) { 2736 case VDIR: 2737 delay = dirdelay; 2738 break; 2739 case VCHR: 2740 delay = metadelay; 2741 break; 2742 default: 2743 delay = filedelay; 2744 } 2745 vn_syncer_add_to_worklist(bo, delay); 2746 } 2747 buf_vlist_add(bp, bo, BX_VNDIRTY); 2748 } else { 2749 buf_vlist_add(bp, bo, BX_VNCLEAN); 2750 2751 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2752 mtx_lock(&sync_mtx); 2753 LIST_REMOVE(bo, bo_synclist); 2754 syncer_worklist_len--; 2755 mtx_unlock(&sync_mtx); 2756 bo->bo_flag &= ~BO_ONWORKLST; 2757 } 2758 } 2759 #ifdef INVARIANTS 2760 bv = &bo->bo_clean; 2761 bp = TAILQ_FIRST(&bv->bv_hd); 2762 KASSERT(bp == NULL || bp->b_bufobj == bo, 2763 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2764 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2765 KASSERT(bp == NULL || bp->b_bufobj == bo, 2766 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2767 bv = &bo->bo_dirty; 2768 bp = TAILQ_FIRST(&bv->bv_hd); 2769 KASSERT(bp == NULL || bp->b_bufobj == bo, 2770 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2771 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2772 KASSERT(bp == NULL || bp->b_bufobj == bo, 2773 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2774 #endif 2775 BO_UNLOCK(bo); 2776 } 2777 2778 static void 2779 v_init_counters(struct vnode *vp) 2780 { 2781 2782 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 2783 vp, ("%s called for an initialized vnode", __FUNCTION__)); 2784 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 2785 2786 refcount_init(&vp->v_holdcnt, 1); 2787 refcount_init(&vp->v_usecount, 1); 2788 } 2789 2790 /* 2791 * Increment si_usecount of the associated device, if any. 2792 */ 2793 static void 2794 v_incr_devcount(struct vnode *vp) 2795 { 2796 2797 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2798 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2799 dev_lock(); 2800 vp->v_rdev->si_usecount++; 2801 dev_unlock(); 2802 } 2803 } 2804 2805 /* 2806 * Decrement si_usecount of the associated device, if any. 2807 */ 2808 static void 2809 v_decr_devcount(struct vnode *vp) 2810 { 2811 2812 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2813 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2814 dev_lock(); 2815 vp->v_rdev->si_usecount--; 2816 dev_unlock(); 2817 } 2818 } 2819 2820 /* 2821 * Grab a particular vnode from the free list, increment its 2822 * reference count and lock it. VIRF_DOOMED is set if the vnode 2823 * is being destroyed. Only callers who specify LK_RETRY will 2824 * see doomed vnodes. If inactive processing was delayed in 2825 * vput try to do it here. 2826 * 2827 * Both holdcnt and usecount can be manipulated using atomics without holding 2828 * any locks except in these cases which require the vnode interlock: 2829 * holdcnt: 1->0 and 0->1 2830 * usecount: 0->1 2831 * 2832 * usecount is permitted to transition 1->0 without the interlock because 2833 * vnode is kept live by holdcnt. 2834 */ 2835 static enum vgetstate __always_inline 2836 _vget_prep(struct vnode *vp, bool interlock) 2837 { 2838 enum vgetstate vs; 2839 2840 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2841 vs = VGET_USECOUNT; 2842 } else { 2843 if (interlock) 2844 vholdl(vp); 2845 else 2846 vhold(vp); 2847 vs = VGET_HOLDCNT; 2848 } 2849 return (vs); 2850 } 2851 2852 enum vgetstate 2853 vget_prep(struct vnode *vp) 2854 { 2855 2856 return (_vget_prep(vp, false)); 2857 } 2858 2859 int 2860 vget(struct vnode *vp, int flags, struct thread *td) 2861 { 2862 enum vgetstate vs; 2863 2864 MPASS(td == curthread); 2865 2866 vs = _vget_prep(vp, (flags & LK_INTERLOCK) != 0); 2867 return (vget_finish(vp, flags, vs)); 2868 } 2869 2870 int 2871 vget_finish(struct vnode *vp, int flags, enum vgetstate vs) 2872 { 2873 int error, oweinact; 2874 2875 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 2876 ("%s: invalid lock operation", __func__)); 2877 2878 if ((flags & LK_INTERLOCK) != 0) 2879 ASSERT_VI_LOCKED(vp, __func__); 2880 else 2881 ASSERT_VI_UNLOCKED(vp, __func__); 2882 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 2883 if (vs == VGET_USECOUNT) { 2884 VNASSERT(vp->v_usecount > 0, vp, 2885 ("%s: vnode without usecount when VGET_USECOUNT was passed", 2886 __func__)); 2887 } 2888 2889 if ((error = vn_lock(vp, flags)) != 0) { 2890 if (vs == VGET_USECOUNT) 2891 vrele(vp); 2892 else 2893 vdrop(vp); 2894 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 2895 vp); 2896 return (error); 2897 } 2898 2899 if (vs == VGET_USECOUNT) { 2900 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2901 ("%s: vnode with usecount and VI_OWEINACT set", __func__)); 2902 return (0); 2903 } 2904 2905 /* 2906 * We hold the vnode. If the usecount is 0 it will be utilized to keep 2907 * the vnode around. Otherwise someone else lended their hold count and 2908 * we have to drop ours. 2909 */ 2910 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2911 #ifdef INVARIANTS 2912 int old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 2913 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 2914 #else 2915 refcount_release(&vp->v_holdcnt); 2916 #endif 2917 VNODE_REFCOUNT_FENCE_ACQ(); 2918 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2919 ("%s: vnode with usecount and VI_OWEINACT set", __func__)); 2920 return (0); 2921 } 2922 2923 /* 2924 * We don't guarantee that any particular close will 2925 * trigger inactive processing so just make a best effort 2926 * here at preventing a reference to a removed file. If 2927 * we don't succeed no harm is done. 2928 * 2929 * Upgrade our holdcnt to a usecount. 2930 */ 2931 VI_LOCK(vp); 2932 /* 2933 * See the previous section. By the time we get here we may find 2934 * ourselves in the same spot. 2935 */ 2936 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2937 #ifdef INVARIANTS 2938 int old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 2939 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 2940 #else 2941 refcount_release(&vp->v_holdcnt); 2942 #endif 2943 VNODE_REFCOUNT_FENCE_ACQ(); 2944 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2945 ("%s: vnode with usecount and VI_OWEINACT set", 2946 __func__)); 2947 VI_UNLOCK(vp); 2948 return (0); 2949 } 2950 if ((vp->v_iflag & VI_OWEINACT) == 0) { 2951 oweinact = 0; 2952 } else { 2953 oweinact = 1; 2954 vp->v_iflag &= ~VI_OWEINACT; 2955 VNODE_REFCOUNT_FENCE_REL(); 2956 } 2957 v_incr_devcount(vp); 2958 refcount_acquire(&vp->v_usecount); 2959 if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE && 2960 (flags & LK_NOWAIT) == 0) 2961 vinactive(vp); 2962 VI_UNLOCK(vp); 2963 return (0); 2964 } 2965 2966 /* 2967 * Increase the reference (use) and hold count of a vnode. 2968 * This will also remove the vnode from the free list if it is presently free. 2969 */ 2970 void 2971 vref(struct vnode *vp) 2972 { 2973 2974 ASSERT_VI_UNLOCKED(vp, __func__); 2975 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2976 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2977 VNODE_REFCOUNT_FENCE_ACQ(); 2978 VNASSERT(vp->v_holdcnt > 0, vp, 2979 ("%s: active vnode not held", __func__)); 2980 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2981 ("%s: vnode with usecount and VI_OWEINACT set", __func__)); 2982 return; 2983 } 2984 VI_LOCK(vp); 2985 vrefl(vp); 2986 VI_UNLOCK(vp); 2987 } 2988 2989 void 2990 vrefl(struct vnode *vp) 2991 { 2992 2993 ASSERT_VI_LOCKED(vp, __func__); 2994 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2995 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2996 VNODE_REFCOUNT_FENCE_ACQ(); 2997 VNASSERT(vp->v_holdcnt > 0, vp, 2998 ("%s: active vnode not held", __func__)); 2999 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 3000 ("%s: vnode with usecount and VI_OWEINACT set", __func__)); 3001 return; 3002 } 3003 vholdl(vp); 3004 if ((vp->v_iflag & VI_OWEINACT) != 0) { 3005 vp->v_iflag &= ~VI_OWEINACT; 3006 VNODE_REFCOUNT_FENCE_REL(); 3007 } 3008 v_incr_devcount(vp); 3009 refcount_acquire(&vp->v_usecount); 3010 } 3011 3012 void 3013 vrefact(struct vnode *vp) 3014 { 3015 3016 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3017 #ifdef INVARIANTS 3018 int old = atomic_fetchadd_int(&vp->v_usecount, 1); 3019 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 3020 #else 3021 refcount_acquire(&vp->v_usecount); 3022 #endif 3023 } 3024 3025 /* 3026 * Return reference count of a vnode. 3027 * 3028 * The results of this call are only guaranteed when some mechanism is used to 3029 * stop other processes from gaining references to the vnode. This may be the 3030 * case if the caller holds the only reference. This is also useful when stale 3031 * data is acceptable as race conditions may be accounted for by some other 3032 * means. 3033 */ 3034 int 3035 vrefcnt(struct vnode *vp) 3036 { 3037 3038 return (vp->v_usecount); 3039 } 3040 3041 static void 3042 vdefer_inactive(struct vnode *vp) 3043 { 3044 3045 ASSERT_VI_LOCKED(vp, __func__); 3046 VNASSERT(vp->v_iflag & VI_OWEINACT, vp, 3047 ("%s: vnode without VI_OWEINACT", __func__)); 3048 if (VN_IS_DOOMED(vp)) { 3049 vdropl(vp); 3050 return; 3051 } 3052 if (vp->v_iflag & VI_DEFINACT) { 3053 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 3054 vdropl(vp); 3055 return; 3056 } 3057 vp->v_iflag |= VI_DEFINACT; 3058 VI_UNLOCK(vp); 3059 counter_u64_add(deferred_inact, 1); 3060 } 3061 3062 static void 3063 vdefer_inactive_cond(struct vnode *vp) 3064 { 3065 3066 VI_LOCK(vp); 3067 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 3068 if ((vp->v_iflag & VI_OWEINACT) == 0) { 3069 vdropl(vp); 3070 return; 3071 } 3072 vdefer_inactive(vp); 3073 } 3074 3075 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF }; 3076 3077 /* 3078 * Decrement the use and hold counts for a vnode. 3079 * 3080 * See an explanation near vget() as to why atomic operation is safe. 3081 */ 3082 static void 3083 vputx(struct vnode *vp, enum vputx_op func) 3084 { 3085 int error; 3086 3087 KASSERT(vp != NULL, ("vputx: null vp")); 3088 if (func == VPUTX_VUNREF) 3089 ASSERT_VOP_LOCKED(vp, "vunref"); 3090 ASSERT_VI_UNLOCKED(vp, __func__); 3091 VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp, 3092 ("%s: wrong ref counts", __func__)); 3093 3094 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3095 3096 /* 3097 * We want to hold the vnode until the inactive finishes to 3098 * prevent vgone() races. We drop the use count here and the 3099 * hold count below when we're done. 3100 * 3101 * If we release the last usecount we take ownership of the hold 3102 * count which provides liveness of the vnode, in which case we 3103 * have to vdrop. 3104 */ 3105 if (!refcount_release(&vp->v_usecount)) 3106 return; 3107 VI_LOCK(vp); 3108 v_decr_devcount(vp); 3109 /* 3110 * By the time we got here someone else might have transitioned 3111 * the count back to > 0. 3112 */ 3113 if (vp->v_usecount > 0) { 3114 vdropl(vp); 3115 return; 3116 } 3117 if (vp->v_iflag & VI_DOINGINACT) { 3118 vdropl(vp); 3119 return; 3120 } 3121 3122 /* 3123 * Check if the fs wants to perform inactive processing. Note we 3124 * may be only holding the interlock, in which case it is possible 3125 * someone else called vgone on the vnode and ->v_data is now NULL. 3126 * Since vgone performs inactive on its own there is nothing to do 3127 * here but to drop our hold count. 3128 */ 3129 if (__predict_false(VN_IS_DOOMED(vp)) || 3130 VOP_NEED_INACTIVE(vp) == 0) { 3131 vdropl(vp); 3132 return; 3133 } 3134 3135 /* 3136 * We must call VOP_INACTIVE with the node locked. Mark 3137 * as VI_DOINGINACT to avoid recursion. 3138 */ 3139 vp->v_iflag |= VI_OWEINACT; 3140 switch (func) { 3141 case VPUTX_VRELE: 3142 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3143 VI_LOCK(vp); 3144 break; 3145 case VPUTX_VPUT: 3146 error = VOP_LOCK(vp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT); 3147 VI_LOCK(vp); 3148 break; 3149 case VPUTX_VUNREF: 3150 error = 0; 3151 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3152 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3153 VI_LOCK(vp); 3154 } 3155 break; 3156 } 3157 VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp, 3158 ("vnode with usecount and VI_OWEINACT set")); 3159 if (error == 0) { 3160 if (vp->v_iflag & VI_OWEINACT) 3161 vinactive(vp); 3162 if (func != VPUTX_VUNREF) 3163 VOP_UNLOCK(vp); 3164 vdropl(vp); 3165 } else if (vp->v_iflag & VI_OWEINACT) { 3166 vdefer_inactive(vp); 3167 } else { 3168 vdropl(vp); 3169 } 3170 } 3171 3172 /* 3173 * Vnode put/release. 3174 * If count drops to zero, call inactive routine and return to freelist. 3175 */ 3176 void 3177 vrele(struct vnode *vp) 3178 { 3179 3180 vputx(vp, VPUTX_VRELE); 3181 } 3182 3183 /* 3184 * Release an already locked vnode. This give the same effects as 3185 * unlock+vrele(), but takes less time and avoids releasing and 3186 * re-aquiring the lock (as vrele() acquires the lock internally.) 3187 * 3188 * It is an invariant that all VOP_* calls operate on a held vnode. 3189 * We may be only having an implicit hold stemming from our usecount, 3190 * which we are about to release. If we unlock the vnode afterwards we 3191 * open a time window where someone else dropped the last usecount and 3192 * proceeded to free the vnode before our unlock finished. For this 3193 * reason we unlock the vnode early. This is a little bit wasteful as 3194 * it may be the vnode is exclusively locked and inactive processing is 3195 * needed, in which case we are adding work. 3196 */ 3197 void 3198 vput(struct vnode *vp) 3199 { 3200 3201 VOP_UNLOCK(vp); 3202 vputx(vp, VPUTX_VPUT); 3203 } 3204 3205 /* 3206 * Release an exclusively locked vnode. Do not unlock the vnode lock. 3207 */ 3208 void 3209 vunref(struct vnode *vp) 3210 { 3211 3212 vputx(vp, VPUTX_VUNREF); 3213 } 3214 3215 /* 3216 * Increase the hold count and activate if this is the first reference. 3217 */ 3218 static void 3219 vhold_activate(struct vnode *vp) 3220 { 3221 struct mount *mp; 3222 3223 ASSERT_VI_LOCKED(vp, __func__); 3224 VNASSERT(vp->v_holdcnt == 0, vp, 3225 ("%s: wrong hold count", __func__)); 3226 VNASSERT(vp->v_op != NULL, vp, 3227 ("%s: vnode already reclaimed.", __func__)); 3228 /* 3229 * Remove a vnode from the free list, mark it as in use, 3230 * and put it on the active list. 3231 */ 3232 VNASSERT(vp->v_mount != NULL, vp, 3233 ("_vhold: vnode not on per mount vnode list")); 3234 mp = vp->v_mount; 3235 mtx_lock(&mp->mnt_listmtx); 3236 if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) { 3237 TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist); 3238 mp->mnt_tmpfreevnodelistsize--; 3239 vp->v_mflag &= ~VMP_TMPMNTFREELIST; 3240 } else { 3241 mtx_lock(&vnode_free_list_mtx); 3242 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 3243 freevnodes--; 3244 mtx_unlock(&vnode_free_list_mtx); 3245 } 3246 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 3247 ("Activating already active vnode")); 3248 vp->v_iflag &= ~VI_FREE; 3249 vp->v_iflag |= VI_ACTIVE; 3250 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 3251 mp->mnt_activevnodelistsize++; 3252 mtx_unlock(&mp->mnt_listmtx); 3253 refcount_acquire(&vp->v_holdcnt); 3254 } 3255 3256 void 3257 vhold(struct vnode *vp) 3258 { 3259 3260 ASSERT_VI_UNLOCKED(vp, __func__); 3261 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3262 if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) { 3263 VNODE_REFCOUNT_FENCE_ACQ(); 3264 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 3265 ("vhold: vnode with holdcnt is free")); 3266 return; 3267 } 3268 VI_LOCK(vp); 3269 vholdl(vp); 3270 VI_UNLOCK(vp); 3271 } 3272 3273 void 3274 vholdl(struct vnode *vp) 3275 { 3276 3277 ASSERT_VI_LOCKED(vp, __func__); 3278 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3279 if ((vp->v_iflag & VI_FREE) == 0) { 3280 refcount_acquire(&vp->v_holdcnt); 3281 return; 3282 } 3283 vhold_activate(vp); 3284 } 3285 3286 void 3287 vholdnz(struct vnode *vp) 3288 { 3289 3290 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3291 #ifdef INVARIANTS 3292 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3293 VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old)); 3294 #else 3295 atomic_add_int(&vp->v_holdcnt, 1); 3296 #endif 3297 } 3298 3299 /* 3300 * Drop the hold count of the vnode. If this is the last reference to 3301 * the vnode we place it on the free list unless it has been vgone'd 3302 * (marked VIRF_DOOMED) in which case we will free it. 3303 * 3304 * Because the vnode vm object keeps a hold reference on the vnode if 3305 * there is at least one resident non-cached page, the vnode cannot 3306 * leave the active list without the page cleanup done. 3307 */ 3308 static void 3309 vdrop_deactivate(struct vnode *vp) 3310 { 3311 struct mount *mp; 3312 3313 ASSERT_VI_LOCKED(vp, __func__); 3314 /* 3315 * Mark a vnode as free: remove it from its active list 3316 * and put it up for recycling on the freelist. 3317 */ 3318 VNASSERT(!VN_IS_DOOMED(vp), vp, 3319 ("vdrop: returning doomed vnode")); 3320 VNASSERT(vp->v_op != NULL, vp, 3321 ("vdrop: vnode already reclaimed.")); 3322 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 3323 ("vnode already free")); 3324 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 3325 ("vnode with VI_OWEINACT set")); 3326 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, 3327 ("vnode with VI_DEFINACT set")); 3328 VNASSERT(vp->v_holdcnt == 0, vp, 3329 ("vdrop: freeing when we shouldn't")); 3330 mp = vp->v_mount; 3331 mtx_lock(&mp->mnt_listmtx); 3332 if (vp->v_iflag & VI_ACTIVE) { 3333 vp->v_iflag &= ~VI_ACTIVE; 3334 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist); 3335 mp->mnt_activevnodelistsize--; 3336 } 3337 TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist); 3338 mp->mnt_tmpfreevnodelistsize++; 3339 vp->v_iflag |= VI_FREE; 3340 vp->v_mflag |= VMP_TMPMNTFREELIST; 3341 VI_UNLOCK(vp); 3342 if (mp->mnt_tmpfreevnodelistsize >= mnt_free_list_batch) 3343 vnlru_return_batch_locked(mp); 3344 mtx_unlock(&mp->mnt_listmtx); 3345 } 3346 3347 void 3348 vdrop(struct vnode *vp) 3349 { 3350 3351 ASSERT_VI_UNLOCKED(vp, __func__); 3352 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3353 if (refcount_release_if_not_last(&vp->v_holdcnt)) 3354 return; 3355 VI_LOCK(vp); 3356 vdropl(vp); 3357 } 3358 3359 void 3360 vdropl(struct vnode *vp) 3361 { 3362 3363 ASSERT_VI_LOCKED(vp, __func__); 3364 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3365 if (!refcount_release(&vp->v_holdcnt)) { 3366 VI_UNLOCK(vp); 3367 return; 3368 } 3369 if (VN_IS_DOOMED(vp)) { 3370 freevnode(vp); 3371 return; 3372 } 3373 vdrop_deactivate(vp); 3374 } 3375 3376 /* 3377 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 3378 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 3379 * OWEINACT tracks whether a vnode missed a call to inactive due to a 3380 * failed lock upgrade. 3381 */ 3382 void 3383 vinactive(struct vnode *vp) 3384 { 3385 struct vm_object *obj; 3386 3387 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3388 ASSERT_VI_LOCKED(vp, "vinactive"); 3389 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 3390 ("vinactive: recursed on VI_DOINGINACT")); 3391 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3392 vp->v_iflag |= VI_DOINGINACT; 3393 vp->v_iflag &= ~VI_OWEINACT; 3394 VI_UNLOCK(vp); 3395 /* 3396 * Before moving off the active list, we must be sure that any 3397 * modified pages are converted into the vnode's dirty 3398 * buffers, since these will no longer be checked once the 3399 * vnode is on the inactive list. 3400 * 3401 * The write-out of the dirty pages is asynchronous. At the 3402 * point that VOP_INACTIVE() is called, there could still be 3403 * pending I/O and dirty pages in the object. 3404 */ 3405 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 3406 vm_object_mightbedirty(obj)) { 3407 VM_OBJECT_WLOCK(obj); 3408 vm_object_page_clean(obj, 0, 0, 0); 3409 VM_OBJECT_WUNLOCK(obj); 3410 } 3411 VOP_INACTIVE(vp, curthread); 3412 VI_LOCK(vp); 3413 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 3414 ("vinactive: lost VI_DOINGINACT")); 3415 vp->v_iflag &= ~VI_DOINGINACT; 3416 } 3417 3418 /* 3419 * Remove any vnodes in the vnode table belonging to mount point mp. 3420 * 3421 * If FORCECLOSE is not specified, there should not be any active ones, 3422 * return error if any are found (nb: this is a user error, not a 3423 * system error). If FORCECLOSE is specified, detach any active vnodes 3424 * that are found. 3425 * 3426 * If WRITECLOSE is set, only flush out regular file vnodes open for 3427 * writing. 3428 * 3429 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 3430 * 3431 * `rootrefs' specifies the base reference count for the root vnode 3432 * of this filesystem. The root vnode is considered busy if its 3433 * v_usecount exceeds this value. On a successful return, vflush(, td) 3434 * will call vrele() on the root vnode exactly rootrefs times. 3435 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 3436 * be zero. 3437 */ 3438 #ifdef DIAGNOSTIC 3439 static int busyprt = 0; /* print out busy vnodes */ 3440 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 3441 #endif 3442 3443 int 3444 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 3445 { 3446 struct vnode *vp, *mvp, *rootvp = NULL; 3447 struct vattr vattr; 3448 int busy = 0, error; 3449 3450 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 3451 rootrefs, flags); 3452 if (rootrefs > 0) { 3453 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 3454 ("vflush: bad args")); 3455 /* 3456 * Get the filesystem root vnode. We can vput() it 3457 * immediately, since with rootrefs > 0, it won't go away. 3458 */ 3459 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 3460 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 3461 __func__, error); 3462 return (error); 3463 } 3464 vput(rootvp); 3465 } 3466 loop: 3467 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3468 vholdl(vp); 3469 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 3470 if (error) { 3471 vdrop(vp); 3472 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3473 goto loop; 3474 } 3475 /* 3476 * Skip over a vnodes marked VV_SYSTEM. 3477 */ 3478 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 3479 VOP_UNLOCK(vp); 3480 vdrop(vp); 3481 continue; 3482 } 3483 /* 3484 * If WRITECLOSE is set, flush out unlinked but still open 3485 * files (even if open only for reading) and regular file 3486 * vnodes open for writing. 3487 */ 3488 if (flags & WRITECLOSE) { 3489 if (vp->v_object != NULL) { 3490 VM_OBJECT_WLOCK(vp->v_object); 3491 vm_object_page_clean(vp->v_object, 0, 0, 0); 3492 VM_OBJECT_WUNLOCK(vp->v_object); 3493 } 3494 error = VOP_FSYNC(vp, MNT_WAIT, td); 3495 if (error != 0) { 3496 VOP_UNLOCK(vp); 3497 vdrop(vp); 3498 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3499 return (error); 3500 } 3501 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 3502 VI_LOCK(vp); 3503 3504 if ((vp->v_type == VNON || 3505 (error == 0 && vattr.va_nlink > 0)) && 3506 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 3507 VOP_UNLOCK(vp); 3508 vdropl(vp); 3509 continue; 3510 } 3511 } else 3512 VI_LOCK(vp); 3513 /* 3514 * With v_usecount == 0, all we need to do is clear out the 3515 * vnode data structures and we are done. 3516 * 3517 * If FORCECLOSE is set, forcibly close the vnode. 3518 */ 3519 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 3520 vgonel(vp); 3521 } else { 3522 busy++; 3523 #ifdef DIAGNOSTIC 3524 if (busyprt) 3525 vn_printf(vp, "vflush: busy vnode "); 3526 #endif 3527 } 3528 VOP_UNLOCK(vp); 3529 vdropl(vp); 3530 } 3531 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 3532 /* 3533 * If just the root vnode is busy, and if its refcount 3534 * is equal to `rootrefs', then go ahead and kill it. 3535 */ 3536 VI_LOCK(rootvp); 3537 KASSERT(busy > 0, ("vflush: not busy")); 3538 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 3539 ("vflush: usecount %d < rootrefs %d", 3540 rootvp->v_usecount, rootrefs)); 3541 if (busy == 1 && rootvp->v_usecount == rootrefs) { 3542 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 3543 vgone(rootvp); 3544 VOP_UNLOCK(rootvp); 3545 busy = 0; 3546 } else 3547 VI_UNLOCK(rootvp); 3548 } 3549 if (busy) { 3550 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 3551 busy); 3552 return (EBUSY); 3553 } 3554 for (; rootrefs > 0; rootrefs--) 3555 vrele(rootvp); 3556 return (0); 3557 } 3558 3559 /* 3560 * Recycle an unused vnode to the front of the free list. 3561 */ 3562 int 3563 vrecycle(struct vnode *vp) 3564 { 3565 int recycled; 3566 3567 VI_LOCK(vp); 3568 recycled = vrecyclel(vp); 3569 VI_UNLOCK(vp); 3570 return (recycled); 3571 } 3572 3573 /* 3574 * vrecycle, with the vp interlock held. 3575 */ 3576 int 3577 vrecyclel(struct vnode *vp) 3578 { 3579 int recycled; 3580 3581 ASSERT_VOP_ELOCKED(vp, __func__); 3582 ASSERT_VI_LOCKED(vp, __func__); 3583 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3584 recycled = 0; 3585 if (vp->v_usecount == 0) { 3586 recycled = 1; 3587 vgonel(vp); 3588 } 3589 return (recycled); 3590 } 3591 3592 /* 3593 * Eliminate all activity associated with a vnode 3594 * in preparation for reuse. 3595 */ 3596 void 3597 vgone(struct vnode *vp) 3598 { 3599 VI_LOCK(vp); 3600 vgonel(vp); 3601 VI_UNLOCK(vp); 3602 } 3603 3604 static void 3605 notify_lowervp_vfs_dummy(struct mount *mp __unused, 3606 struct vnode *lowervp __unused) 3607 { 3608 } 3609 3610 /* 3611 * Notify upper mounts about reclaimed or unlinked vnode. 3612 */ 3613 void 3614 vfs_notify_upper(struct vnode *vp, int event) 3615 { 3616 static struct vfsops vgonel_vfsops = { 3617 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, 3618 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, 3619 }; 3620 struct mount *mp, *ump, *mmp; 3621 3622 mp = vp->v_mount; 3623 if (mp == NULL) 3624 return; 3625 if (TAILQ_EMPTY(&mp->mnt_uppers)) 3626 return; 3627 3628 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); 3629 mmp->mnt_op = &vgonel_vfsops; 3630 mmp->mnt_kern_flag |= MNTK_MARKER; 3631 MNT_ILOCK(mp); 3632 mp->mnt_kern_flag |= MNTK_VGONE_UPPER; 3633 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { 3634 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { 3635 ump = TAILQ_NEXT(ump, mnt_upper_link); 3636 continue; 3637 } 3638 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); 3639 MNT_IUNLOCK(mp); 3640 switch (event) { 3641 case VFS_NOTIFY_UPPER_RECLAIM: 3642 VFS_RECLAIM_LOWERVP(ump, vp); 3643 break; 3644 case VFS_NOTIFY_UPPER_UNLINK: 3645 VFS_UNLINK_LOWERVP(ump, vp); 3646 break; 3647 default: 3648 KASSERT(0, ("invalid event %d", event)); 3649 break; 3650 } 3651 MNT_ILOCK(mp); 3652 ump = TAILQ_NEXT(mmp, mnt_upper_link); 3653 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); 3654 } 3655 free(mmp, M_TEMP); 3656 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; 3657 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { 3658 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; 3659 wakeup(&mp->mnt_uppers); 3660 } 3661 MNT_IUNLOCK(mp); 3662 } 3663 3664 /* 3665 * vgone, with the vp interlock held. 3666 */ 3667 static void 3668 vgonel(struct vnode *vp) 3669 { 3670 struct thread *td; 3671 struct mount *mp; 3672 vm_object_t object; 3673 bool active, oweinact; 3674 3675 ASSERT_VOP_ELOCKED(vp, "vgonel"); 3676 ASSERT_VI_LOCKED(vp, "vgonel"); 3677 VNASSERT(vp->v_holdcnt, vp, 3678 ("vgonel: vp %p has no reference.", vp)); 3679 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3680 td = curthread; 3681 3682 /* 3683 * Don't vgonel if we're already doomed. 3684 */ 3685 if (vp->v_irflag & VIRF_DOOMED) 3686 return; 3687 vp->v_irflag |= VIRF_DOOMED; 3688 3689 /* 3690 * Check to see if the vnode is in use. If so, we have to call 3691 * VOP_CLOSE() and VOP_INACTIVE(). 3692 */ 3693 active = vp->v_usecount > 0; 3694 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 3695 /* 3696 * If we need to do inactive VI_OWEINACT will be set. 3697 */ 3698 if (vp->v_iflag & VI_DEFINACT) { 3699 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 3700 vp->v_iflag &= ~VI_DEFINACT; 3701 vdropl(vp); 3702 } else { 3703 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 3704 VI_UNLOCK(vp); 3705 } 3706 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 3707 3708 /* 3709 * If purging an active vnode, it must be closed and 3710 * deactivated before being reclaimed. 3711 */ 3712 if (active) 3713 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 3714 if (oweinact || active) { 3715 VI_LOCK(vp); 3716 if ((vp->v_iflag & VI_DOINGINACT) == 0) 3717 vinactive(vp); 3718 VI_UNLOCK(vp); 3719 } 3720 if (vp->v_type == VSOCK) 3721 vfs_unp_reclaim(vp); 3722 3723 /* 3724 * Clean out any buffers associated with the vnode. 3725 * If the flush fails, just toss the buffers. 3726 */ 3727 mp = NULL; 3728 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 3729 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 3730 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 3731 while (vinvalbuf(vp, 0, 0, 0) != 0) 3732 ; 3733 } 3734 3735 BO_LOCK(&vp->v_bufobj); 3736 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 3737 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 3738 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 3739 vp->v_bufobj.bo_clean.bv_cnt == 0, 3740 ("vp %p bufobj not invalidated", vp)); 3741 3742 /* 3743 * For VMIO bufobj, BO_DEAD is set later, or in 3744 * vm_object_terminate() after the object's page queue is 3745 * flushed. 3746 */ 3747 object = vp->v_bufobj.bo_object; 3748 if (object == NULL) 3749 vp->v_bufobj.bo_flag |= BO_DEAD; 3750 BO_UNLOCK(&vp->v_bufobj); 3751 3752 /* 3753 * Handle the VM part. Tmpfs handles v_object on its own (the 3754 * OBJT_VNODE check). Nullfs or other bypassing filesystems 3755 * should not touch the object borrowed from the lower vnode 3756 * (the handle check). 3757 */ 3758 if (object != NULL && object->type == OBJT_VNODE && 3759 object->handle == vp) 3760 vnode_destroy_vobject(vp); 3761 3762 /* 3763 * Reclaim the vnode. 3764 */ 3765 if (VOP_RECLAIM(vp, td)) 3766 panic("vgone: cannot reclaim"); 3767 if (mp != NULL) 3768 vn_finished_secondary_write(mp); 3769 VNASSERT(vp->v_object == NULL, vp, 3770 ("vop_reclaim left v_object vp=%p", vp)); 3771 /* 3772 * Clear the advisory locks and wake up waiting threads. 3773 */ 3774 (void)VOP_ADVLOCKPURGE(vp); 3775 vp->v_lockf = NULL; 3776 /* 3777 * Delete from old mount point vnode list. 3778 */ 3779 delmntque(vp); 3780 cache_purge(vp); 3781 /* 3782 * Done with purge, reset to the standard lock and invalidate 3783 * the vnode. 3784 */ 3785 VI_LOCK(vp); 3786 vp->v_vnlock = &vp->v_lock; 3787 vp->v_op = &dead_vnodeops; 3788 vp->v_type = VBAD; 3789 } 3790 3791 /* 3792 * Calculate the total number of references to a special device. 3793 */ 3794 int 3795 vcount(struct vnode *vp) 3796 { 3797 int count; 3798 3799 dev_lock(); 3800 count = vp->v_rdev->si_usecount; 3801 dev_unlock(); 3802 return (count); 3803 } 3804 3805 /* 3806 * Print out a description of a vnode. 3807 */ 3808 static char *typename[] = 3809 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 3810 "VMARKER"}; 3811 3812 void 3813 vn_printf(struct vnode *vp, const char *fmt, ...) 3814 { 3815 va_list ap; 3816 char buf[256], buf2[16]; 3817 u_long flags; 3818 3819 va_start(ap, fmt); 3820 vprintf(fmt, ap); 3821 va_end(ap); 3822 printf("%p: ", (void *)vp); 3823 printf("type %s\n", typename[vp->v_type]); 3824 printf(" usecount %d, writecount %d, refcount %d", 3825 vp->v_usecount, vp->v_writecount, vp->v_holdcnt); 3826 switch (vp->v_type) { 3827 case VDIR: 3828 printf(" mountedhere %p\n", vp->v_mountedhere); 3829 break; 3830 case VCHR: 3831 printf(" rdev %p\n", vp->v_rdev); 3832 break; 3833 case VSOCK: 3834 printf(" socket %p\n", vp->v_unpcb); 3835 break; 3836 case VFIFO: 3837 printf(" fifoinfo %p\n", vp->v_fifoinfo); 3838 break; 3839 default: 3840 printf("\n"); 3841 break; 3842 } 3843 buf[0] = '\0'; 3844 buf[1] = '\0'; 3845 if (vp->v_irflag & VIRF_DOOMED) 3846 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 3847 flags = vp->v_irflag & ~(VIRF_DOOMED); 3848 if (flags != 0) { 3849 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 3850 strlcat(buf, buf2, sizeof(buf)); 3851 } 3852 if (vp->v_vflag & VV_ROOT) 3853 strlcat(buf, "|VV_ROOT", sizeof(buf)); 3854 if (vp->v_vflag & VV_ISTTY) 3855 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 3856 if (vp->v_vflag & VV_NOSYNC) 3857 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 3858 if (vp->v_vflag & VV_ETERNALDEV) 3859 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 3860 if (vp->v_vflag & VV_CACHEDLABEL) 3861 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 3862 if (vp->v_vflag & VV_VMSIZEVNLOCK) 3863 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 3864 if (vp->v_vflag & VV_COPYONWRITE) 3865 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 3866 if (vp->v_vflag & VV_SYSTEM) 3867 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 3868 if (vp->v_vflag & VV_PROCDEP) 3869 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 3870 if (vp->v_vflag & VV_NOKNOTE) 3871 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 3872 if (vp->v_vflag & VV_DELETED) 3873 strlcat(buf, "|VV_DELETED", sizeof(buf)); 3874 if (vp->v_vflag & VV_MD) 3875 strlcat(buf, "|VV_MD", sizeof(buf)); 3876 if (vp->v_vflag & VV_FORCEINSMQ) 3877 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 3878 if (vp->v_vflag & VV_READLINK) 3879 strlcat(buf, "|VV_READLINK", sizeof(buf)); 3880 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 3881 VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 3882 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); 3883 if (flags != 0) { 3884 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 3885 strlcat(buf, buf2, sizeof(buf)); 3886 } 3887 if (vp->v_iflag & VI_TEXT_REF) 3888 strlcat(buf, "|VI_TEXT_REF", sizeof(buf)); 3889 if (vp->v_iflag & VI_MOUNT) 3890 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 3891 if (vp->v_iflag & VI_FREE) 3892 strlcat(buf, "|VI_FREE", sizeof(buf)); 3893 if (vp->v_iflag & VI_ACTIVE) 3894 strlcat(buf, "|VI_ACTIVE", sizeof(buf)); 3895 if (vp->v_iflag & VI_DOINGINACT) 3896 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 3897 if (vp->v_iflag & VI_OWEINACT) 3898 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 3899 if (vp->v_iflag & VI_DEFINACT) 3900 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 3901 flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_FREE | VI_ACTIVE | 3902 VI_DOINGINACT | VI_OWEINACT | VI_DEFINACT); 3903 if (flags != 0) { 3904 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 3905 strlcat(buf, buf2, sizeof(buf)); 3906 } 3907 if (vp->v_mflag & VMP_TMPMNTFREELIST) 3908 strlcat(buf, "|VMP_TMPMNTFREELIST", sizeof(buf)); 3909 flags = vp->v_mflag & ~(VMP_TMPMNTFREELIST); 3910 if (flags != 0) { 3911 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 3912 strlcat(buf, buf2, sizeof(buf)); 3913 } 3914 printf(" flags (%s)\n", buf + 1); 3915 if (mtx_owned(VI_MTX(vp))) 3916 printf(" VI_LOCKed"); 3917 if (vp->v_object != NULL) 3918 printf(" v_object %p ref %d pages %d " 3919 "cleanbuf %d dirtybuf %d\n", 3920 vp->v_object, vp->v_object->ref_count, 3921 vp->v_object->resident_page_count, 3922 vp->v_bufobj.bo_clean.bv_cnt, 3923 vp->v_bufobj.bo_dirty.bv_cnt); 3924 printf(" "); 3925 lockmgr_printinfo(vp->v_vnlock); 3926 if (vp->v_data != NULL) 3927 VOP_PRINT(vp); 3928 } 3929 3930 #ifdef DDB 3931 /* 3932 * List all of the locked vnodes in the system. 3933 * Called when debugging the kernel. 3934 */ 3935 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 3936 { 3937 struct mount *mp; 3938 struct vnode *vp; 3939 3940 /* 3941 * Note: because this is DDB, we can't obey the locking semantics 3942 * for these structures, which means we could catch an inconsistent 3943 * state and dereference a nasty pointer. Not much to be done 3944 * about that. 3945 */ 3946 db_printf("Locked vnodes\n"); 3947 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3948 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3949 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 3950 vn_printf(vp, "vnode "); 3951 } 3952 } 3953 } 3954 3955 /* 3956 * Show details about the given vnode. 3957 */ 3958 DB_SHOW_COMMAND(vnode, db_show_vnode) 3959 { 3960 struct vnode *vp; 3961 3962 if (!have_addr) 3963 return; 3964 vp = (struct vnode *)addr; 3965 vn_printf(vp, "vnode "); 3966 } 3967 3968 /* 3969 * Show details about the given mount point. 3970 */ 3971 DB_SHOW_COMMAND(mount, db_show_mount) 3972 { 3973 struct mount *mp; 3974 struct vfsopt *opt; 3975 struct statfs *sp; 3976 struct vnode *vp; 3977 char buf[512]; 3978 uint64_t mflags; 3979 u_int flags; 3980 3981 if (!have_addr) { 3982 /* No address given, print short info about all mount points. */ 3983 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3984 db_printf("%p %s on %s (%s)\n", mp, 3985 mp->mnt_stat.f_mntfromname, 3986 mp->mnt_stat.f_mntonname, 3987 mp->mnt_stat.f_fstypename); 3988 if (db_pager_quit) 3989 break; 3990 } 3991 db_printf("\nMore info: show mount <addr>\n"); 3992 return; 3993 } 3994 3995 mp = (struct mount *)addr; 3996 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 3997 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 3998 3999 buf[0] = '\0'; 4000 mflags = mp->mnt_flag; 4001 #define MNT_FLAG(flag) do { \ 4002 if (mflags & (flag)) { \ 4003 if (buf[0] != '\0') \ 4004 strlcat(buf, ", ", sizeof(buf)); \ 4005 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4006 mflags &= ~(flag); \ 4007 } \ 4008 } while (0) 4009 MNT_FLAG(MNT_RDONLY); 4010 MNT_FLAG(MNT_SYNCHRONOUS); 4011 MNT_FLAG(MNT_NOEXEC); 4012 MNT_FLAG(MNT_NOSUID); 4013 MNT_FLAG(MNT_NFS4ACLS); 4014 MNT_FLAG(MNT_UNION); 4015 MNT_FLAG(MNT_ASYNC); 4016 MNT_FLAG(MNT_SUIDDIR); 4017 MNT_FLAG(MNT_SOFTDEP); 4018 MNT_FLAG(MNT_NOSYMFOLLOW); 4019 MNT_FLAG(MNT_GJOURNAL); 4020 MNT_FLAG(MNT_MULTILABEL); 4021 MNT_FLAG(MNT_ACLS); 4022 MNT_FLAG(MNT_NOATIME); 4023 MNT_FLAG(MNT_NOCLUSTERR); 4024 MNT_FLAG(MNT_NOCLUSTERW); 4025 MNT_FLAG(MNT_SUJ); 4026 MNT_FLAG(MNT_EXRDONLY); 4027 MNT_FLAG(MNT_EXPORTED); 4028 MNT_FLAG(MNT_DEFEXPORTED); 4029 MNT_FLAG(MNT_EXPORTANON); 4030 MNT_FLAG(MNT_EXKERB); 4031 MNT_FLAG(MNT_EXPUBLIC); 4032 MNT_FLAG(MNT_LOCAL); 4033 MNT_FLAG(MNT_QUOTA); 4034 MNT_FLAG(MNT_ROOTFS); 4035 MNT_FLAG(MNT_USER); 4036 MNT_FLAG(MNT_IGNORE); 4037 MNT_FLAG(MNT_UPDATE); 4038 MNT_FLAG(MNT_DELEXPORT); 4039 MNT_FLAG(MNT_RELOAD); 4040 MNT_FLAG(MNT_FORCE); 4041 MNT_FLAG(MNT_SNAPSHOT); 4042 MNT_FLAG(MNT_BYFSID); 4043 #undef MNT_FLAG 4044 if (mflags != 0) { 4045 if (buf[0] != '\0') 4046 strlcat(buf, ", ", sizeof(buf)); 4047 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4048 "0x%016jx", mflags); 4049 } 4050 db_printf(" mnt_flag = %s\n", buf); 4051 4052 buf[0] = '\0'; 4053 flags = mp->mnt_kern_flag; 4054 #define MNT_KERN_FLAG(flag) do { \ 4055 if (flags & (flag)) { \ 4056 if (buf[0] != '\0') \ 4057 strlcat(buf, ", ", sizeof(buf)); \ 4058 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4059 flags &= ~(flag); \ 4060 } \ 4061 } while (0) 4062 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4063 MNT_KERN_FLAG(MNTK_ASYNC); 4064 MNT_KERN_FLAG(MNTK_SOFTDEP); 4065 MNT_KERN_FLAG(MNTK_DRAINING); 4066 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4067 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4068 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4069 MNT_KERN_FLAG(MNTK_NO_IOPF); 4070 MNT_KERN_FLAG(MNTK_VGONE_UPPER); 4071 MNT_KERN_FLAG(MNTK_VGONE_WAITER); 4072 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); 4073 MNT_KERN_FLAG(MNTK_MARKER); 4074 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4075 MNT_KERN_FLAG(MNTK_NOASYNC); 4076 MNT_KERN_FLAG(MNTK_UNMOUNT); 4077 MNT_KERN_FLAG(MNTK_MWAIT); 4078 MNT_KERN_FLAG(MNTK_SUSPEND); 4079 MNT_KERN_FLAG(MNTK_SUSPEND2); 4080 MNT_KERN_FLAG(MNTK_SUSPENDED); 4081 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4082 MNT_KERN_FLAG(MNTK_NOKNOTE); 4083 #undef MNT_KERN_FLAG 4084 if (flags != 0) { 4085 if (buf[0] != '\0') 4086 strlcat(buf, ", ", sizeof(buf)); 4087 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4088 "0x%08x", flags); 4089 } 4090 db_printf(" mnt_kern_flag = %s\n", buf); 4091 4092 db_printf(" mnt_opt = "); 4093 opt = TAILQ_FIRST(mp->mnt_opt); 4094 if (opt != NULL) { 4095 db_printf("%s", opt->name); 4096 opt = TAILQ_NEXT(opt, link); 4097 while (opt != NULL) { 4098 db_printf(", %s", opt->name); 4099 opt = TAILQ_NEXT(opt, link); 4100 } 4101 } 4102 db_printf("\n"); 4103 4104 sp = &mp->mnt_stat; 4105 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4106 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4107 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4108 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4109 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4110 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4111 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4112 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4113 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4114 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4115 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4116 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4117 4118 db_printf(" mnt_cred = { uid=%u ruid=%u", 4119 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4120 if (jailed(mp->mnt_cred)) 4121 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4122 db_printf(" }\n"); 4123 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4124 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4125 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4126 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4127 db_printf(" mnt_activevnodelistsize = %d\n", 4128 mp->mnt_activevnodelistsize); 4129 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4130 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4131 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 4132 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4133 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4134 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4135 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4136 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4137 db_printf(" mnt_secondary_accwrites = %d\n", 4138 mp->mnt_secondary_accwrites); 4139 db_printf(" mnt_gjprovider = %s\n", 4140 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4141 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4142 4143 db_printf("\n\nList of active vnodes\n"); 4144 TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) { 4145 if (vp->v_type != VMARKER) { 4146 vn_printf(vp, "vnode "); 4147 if (db_pager_quit) 4148 break; 4149 } 4150 } 4151 db_printf("\n\nList of inactive vnodes\n"); 4152 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4153 if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) { 4154 vn_printf(vp, "vnode "); 4155 if (db_pager_quit) 4156 break; 4157 } 4158 } 4159 } 4160 #endif /* DDB */ 4161 4162 /* 4163 * Fill in a struct xvfsconf based on a struct vfsconf. 4164 */ 4165 static int 4166 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4167 { 4168 struct xvfsconf xvfsp; 4169 4170 bzero(&xvfsp, sizeof(xvfsp)); 4171 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4172 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4173 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4174 xvfsp.vfc_flags = vfsp->vfc_flags; 4175 /* 4176 * These are unused in userland, we keep them 4177 * to not break binary compatibility. 4178 */ 4179 xvfsp.vfc_vfsops = NULL; 4180 xvfsp.vfc_next = NULL; 4181 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4182 } 4183 4184 #ifdef COMPAT_FREEBSD32 4185 struct xvfsconf32 { 4186 uint32_t vfc_vfsops; 4187 char vfc_name[MFSNAMELEN]; 4188 int32_t vfc_typenum; 4189 int32_t vfc_refcount; 4190 int32_t vfc_flags; 4191 uint32_t vfc_next; 4192 }; 4193 4194 static int 4195 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4196 { 4197 struct xvfsconf32 xvfsp; 4198 4199 bzero(&xvfsp, sizeof(xvfsp)); 4200 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4201 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4202 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4203 xvfsp.vfc_flags = vfsp->vfc_flags; 4204 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4205 } 4206 #endif 4207 4208 /* 4209 * Top level filesystem related information gathering. 4210 */ 4211 static int 4212 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4213 { 4214 struct vfsconf *vfsp; 4215 int error; 4216 4217 error = 0; 4218 vfsconf_slock(); 4219 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4220 #ifdef COMPAT_FREEBSD32 4221 if (req->flags & SCTL_MASK32) 4222 error = vfsconf2x32(req, vfsp); 4223 else 4224 #endif 4225 error = vfsconf2x(req, vfsp); 4226 if (error) 4227 break; 4228 } 4229 vfsconf_sunlock(); 4230 return (error); 4231 } 4232 4233 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 4234 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 4235 "S,xvfsconf", "List of all configured filesystems"); 4236 4237 #ifndef BURN_BRIDGES 4238 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 4239 4240 static int 4241 vfs_sysctl(SYSCTL_HANDLER_ARGS) 4242 { 4243 int *name = (int *)arg1 - 1; /* XXX */ 4244 u_int namelen = arg2 + 1; /* XXX */ 4245 struct vfsconf *vfsp; 4246 4247 log(LOG_WARNING, "userland calling deprecated sysctl, " 4248 "please rebuild world\n"); 4249 4250 #if 1 || defined(COMPAT_PRELITE2) 4251 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 4252 if (namelen == 1) 4253 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 4254 #endif 4255 4256 switch (name[1]) { 4257 case VFS_MAXTYPENUM: 4258 if (namelen != 2) 4259 return (ENOTDIR); 4260 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 4261 case VFS_CONF: 4262 if (namelen != 3) 4263 return (ENOTDIR); /* overloaded */ 4264 vfsconf_slock(); 4265 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4266 if (vfsp->vfc_typenum == name[2]) 4267 break; 4268 } 4269 vfsconf_sunlock(); 4270 if (vfsp == NULL) 4271 return (EOPNOTSUPP); 4272 #ifdef COMPAT_FREEBSD32 4273 if (req->flags & SCTL_MASK32) 4274 return (vfsconf2x32(req, vfsp)); 4275 else 4276 #endif 4277 return (vfsconf2x(req, vfsp)); 4278 } 4279 return (EOPNOTSUPP); 4280 } 4281 4282 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 4283 CTLFLAG_MPSAFE, vfs_sysctl, 4284 "Generic filesystem"); 4285 4286 #if 1 || defined(COMPAT_PRELITE2) 4287 4288 static int 4289 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 4290 { 4291 int error; 4292 struct vfsconf *vfsp; 4293 struct ovfsconf ovfs; 4294 4295 vfsconf_slock(); 4296 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4297 bzero(&ovfs, sizeof(ovfs)); 4298 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 4299 strcpy(ovfs.vfc_name, vfsp->vfc_name); 4300 ovfs.vfc_index = vfsp->vfc_typenum; 4301 ovfs.vfc_refcount = vfsp->vfc_refcount; 4302 ovfs.vfc_flags = vfsp->vfc_flags; 4303 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 4304 if (error != 0) { 4305 vfsconf_sunlock(); 4306 return (error); 4307 } 4308 } 4309 vfsconf_sunlock(); 4310 return (0); 4311 } 4312 4313 #endif /* 1 || COMPAT_PRELITE2 */ 4314 #endif /* !BURN_BRIDGES */ 4315 4316 #define KINFO_VNODESLOP 10 4317 #ifdef notyet 4318 /* 4319 * Dump vnode list (via sysctl). 4320 */ 4321 /* ARGSUSED */ 4322 static int 4323 sysctl_vnode(SYSCTL_HANDLER_ARGS) 4324 { 4325 struct xvnode *xvn; 4326 struct mount *mp; 4327 struct vnode *vp; 4328 int error, len, n; 4329 4330 /* 4331 * Stale numvnodes access is not fatal here. 4332 */ 4333 req->lock = 0; 4334 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 4335 if (!req->oldptr) 4336 /* Make an estimate */ 4337 return (SYSCTL_OUT(req, 0, len)); 4338 4339 error = sysctl_wire_old_buffer(req, 0); 4340 if (error != 0) 4341 return (error); 4342 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 4343 n = 0; 4344 mtx_lock(&mountlist_mtx); 4345 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4346 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 4347 continue; 4348 MNT_ILOCK(mp); 4349 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4350 if (n == len) 4351 break; 4352 vref(vp); 4353 xvn[n].xv_size = sizeof *xvn; 4354 xvn[n].xv_vnode = vp; 4355 xvn[n].xv_id = 0; /* XXX compat */ 4356 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 4357 XV_COPY(usecount); 4358 XV_COPY(writecount); 4359 XV_COPY(holdcnt); 4360 XV_COPY(mount); 4361 XV_COPY(numoutput); 4362 XV_COPY(type); 4363 #undef XV_COPY 4364 xvn[n].xv_flag = vp->v_vflag; 4365 4366 switch (vp->v_type) { 4367 case VREG: 4368 case VDIR: 4369 case VLNK: 4370 break; 4371 case VBLK: 4372 case VCHR: 4373 if (vp->v_rdev == NULL) { 4374 vrele(vp); 4375 continue; 4376 } 4377 xvn[n].xv_dev = dev2udev(vp->v_rdev); 4378 break; 4379 case VSOCK: 4380 xvn[n].xv_socket = vp->v_socket; 4381 break; 4382 case VFIFO: 4383 xvn[n].xv_fifo = vp->v_fifoinfo; 4384 break; 4385 case VNON: 4386 case VBAD: 4387 default: 4388 /* shouldn't happen? */ 4389 vrele(vp); 4390 continue; 4391 } 4392 vrele(vp); 4393 ++n; 4394 } 4395 MNT_IUNLOCK(mp); 4396 mtx_lock(&mountlist_mtx); 4397 vfs_unbusy(mp); 4398 if (n == len) 4399 break; 4400 } 4401 mtx_unlock(&mountlist_mtx); 4402 4403 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 4404 free(xvn, M_TEMP); 4405 return (error); 4406 } 4407 4408 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | 4409 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", 4410 ""); 4411 #endif 4412 4413 static void 4414 unmount_or_warn(struct mount *mp) 4415 { 4416 int error; 4417 4418 error = dounmount(mp, MNT_FORCE, curthread); 4419 if (error != 0) { 4420 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 4421 if (error == EBUSY) 4422 printf("BUSY)\n"); 4423 else 4424 printf("%d)\n", error); 4425 } 4426 } 4427 4428 /* 4429 * Unmount all filesystems. The list is traversed in reverse order 4430 * of mounting to avoid dependencies. 4431 */ 4432 void 4433 vfs_unmountall(void) 4434 { 4435 struct mount *mp, *tmp; 4436 4437 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 4438 4439 /* 4440 * Since this only runs when rebooting, it is not interlocked. 4441 */ 4442 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 4443 vfs_ref(mp); 4444 4445 /* 4446 * Forcibly unmounting "/dev" before "/" would prevent clean 4447 * unmount of the latter. 4448 */ 4449 if (mp == rootdevmp) 4450 continue; 4451 4452 unmount_or_warn(mp); 4453 } 4454 4455 if (rootdevmp != NULL) 4456 unmount_or_warn(rootdevmp); 4457 } 4458 4459 static void 4460 vfs_deferred_inactive(struct vnode *vp, int lkflags) 4461 { 4462 4463 ASSERT_VI_LOCKED(vp, __func__); 4464 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set")); 4465 if ((vp->v_iflag & VI_OWEINACT) == 0) { 4466 vdropl(vp); 4467 return; 4468 } 4469 if (vn_lock(vp, lkflags) == 0) { 4470 VI_LOCK(vp); 4471 if ((vp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == VI_OWEINACT) 4472 vinactive(vp); 4473 VOP_UNLOCK(vp); 4474 vdropl(vp); 4475 return; 4476 } 4477 vdefer_inactive_cond(vp); 4478 } 4479 4480 static void __noinline 4481 vfs_periodic_inactive(struct mount *mp, int flags) 4482 { 4483 struct vnode *vp, *mvp; 4484 int lkflags; 4485 4486 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4487 if (flags != MNT_WAIT) 4488 lkflags |= LK_NOWAIT; 4489 4490 MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) { 4491 if ((vp->v_iflag & VI_DEFINACT) == 0) { 4492 VI_UNLOCK(vp); 4493 continue; 4494 } 4495 vp->v_iflag &= ~VI_DEFINACT; 4496 vfs_deferred_inactive(vp, lkflags); 4497 } 4498 } 4499 4500 static inline bool 4501 vfs_want_msync(struct vnode *vp) 4502 { 4503 struct vm_object *obj; 4504 4505 if (vp->v_vflag & VV_NOSYNC) 4506 return (false); 4507 obj = vp->v_object; 4508 return (obj != NULL && vm_object_mightbedirty(obj)); 4509 } 4510 4511 static void __noinline 4512 vfs_periodic_msync_inactive(struct mount *mp, int flags) 4513 { 4514 struct vnode *vp, *mvp; 4515 struct vm_object *obj; 4516 struct thread *td; 4517 int lkflags, objflags; 4518 bool seen_defer; 4519 4520 td = curthread; 4521 4522 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4523 if (flags != MNT_WAIT) { 4524 lkflags |= LK_NOWAIT; 4525 objflags = OBJPC_NOSYNC; 4526 } else { 4527 objflags = OBJPC_SYNC; 4528 } 4529 4530 MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) { 4531 seen_defer = false; 4532 if (vp->v_iflag & VI_DEFINACT) { 4533 vp->v_iflag &= ~VI_DEFINACT; 4534 seen_defer = true; 4535 } 4536 if (!vfs_want_msync(vp)) { 4537 if (seen_defer) 4538 vfs_deferred_inactive(vp, lkflags); 4539 else 4540 VI_UNLOCK(vp); 4541 continue; 4542 } 4543 if (vget(vp, lkflags, td) == 0) { 4544 obj = vp->v_object; 4545 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) { 4546 VM_OBJECT_WLOCK(obj); 4547 vm_object_page_clean(obj, 0, 0, objflags); 4548 VM_OBJECT_WUNLOCK(obj); 4549 } 4550 vput(vp); 4551 if (seen_defer) 4552 vdrop(vp); 4553 } else { 4554 if (seen_defer) 4555 vdefer_inactive_cond(vp); 4556 } 4557 } 4558 } 4559 4560 void 4561 vfs_periodic(struct mount *mp, int flags) 4562 { 4563 4564 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 4565 4566 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 4567 vfs_periodic_inactive(mp, flags); 4568 else 4569 vfs_periodic_msync_inactive(mp, flags); 4570 } 4571 4572 static void 4573 destroy_vpollinfo_free(struct vpollinfo *vi) 4574 { 4575 4576 knlist_destroy(&vi->vpi_selinfo.si_note); 4577 mtx_destroy(&vi->vpi_lock); 4578 uma_zfree(vnodepoll_zone, vi); 4579 } 4580 4581 static void 4582 destroy_vpollinfo(struct vpollinfo *vi) 4583 { 4584 4585 knlist_clear(&vi->vpi_selinfo.si_note, 1); 4586 seldrain(&vi->vpi_selinfo); 4587 destroy_vpollinfo_free(vi); 4588 } 4589 4590 /* 4591 * Initialize per-vnode helper structure to hold poll-related state. 4592 */ 4593 void 4594 v_addpollinfo(struct vnode *vp) 4595 { 4596 struct vpollinfo *vi; 4597 4598 if (vp->v_pollinfo != NULL) 4599 return; 4600 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO); 4601 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 4602 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 4603 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); 4604 VI_LOCK(vp); 4605 if (vp->v_pollinfo != NULL) { 4606 VI_UNLOCK(vp); 4607 destroy_vpollinfo_free(vi); 4608 return; 4609 } 4610 vp->v_pollinfo = vi; 4611 VI_UNLOCK(vp); 4612 } 4613 4614 /* 4615 * Record a process's interest in events which might happen to 4616 * a vnode. Because poll uses the historic select-style interface 4617 * internally, this routine serves as both the ``check for any 4618 * pending events'' and the ``record my interest in future events'' 4619 * functions. (These are done together, while the lock is held, 4620 * to avoid race conditions.) 4621 */ 4622 int 4623 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 4624 { 4625 4626 v_addpollinfo(vp); 4627 mtx_lock(&vp->v_pollinfo->vpi_lock); 4628 if (vp->v_pollinfo->vpi_revents & events) { 4629 /* 4630 * This leaves events we are not interested 4631 * in available for the other process which 4632 * which presumably had requested them 4633 * (otherwise they would never have been 4634 * recorded). 4635 */ 4636 events &= vp->v_pollinfo->vpi_revents; 4637 vp->v_pollinfo->vpi_revents &= ~events; 4638 4639 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4640 return (events); 4641 } 4642 vp->v_pollinfo->vpi_events |= events; 4643 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 4644 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4645 return (0); 4646 } 4647 4648 /* 4649 * Routine to create and manage a filesystem syncer vnode. 4650 */ 4651 #define sync_close ((int (*)(struct vop_close_args *))nullop) 4652 static int sync_fsync(struct vop_fsync_args *); 4653 static int sync_inactive(struct vop_inactive_args *); 4654 static int sync_reclaim(struct vop_reclaim_args *); 4655 4656 static struct vop_vector sync_vnodeops = { 4657 .vop_bypass = VOP_EOPNOTSUPP, 4658 .vop_close = sync_close, /* close */ 4659 .vop_fsync = sync_fsync, /* fsync */ 4660 .vop_inactive = sync_inactive, /* inactive */ 4661 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */ 4662 .vop_reclaim = sync_reclaim, /* reclaim */ 4663 .vop_lock1 = vop_stdlock, /* lock */ 4664 .vop_unlock = vop_stdunlock, /* unlock */ 4665 .vop_islocked = vop_stdislocked, /* islocked */ 4666 }; 4667 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 4668 4669 /* 4670 * Create a new filesystem syncer vnode for the specified mount point. 4671 */ 4672 void 4673 vfs_allocate_syncvnode(struct mount *mp) 4674 { 4675 struct vnode *vp; 4676 struct bufobj *bo; 4677 static long start, incr, next; 4678 int error; 4679 4680 /* Allocate a new vnode */ 4681 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 4682 if (error != 0) 4683 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 4684 vp->v_type = VNON; 4685 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4686 vp->v_vflag |= VV_FORCEINSMQ; 4687 error = insmntque(vp, mp); 4688 if (error != 0) 4689 panic("vfs_allocate_syncvnode: insmntque() failed"); 4690 vp->v_vflag &= ~VV_FORCEINSMQ; 4691 VOP_UNLOCK(vp); 4692 /* 4693 * Place the vnode onto the syncer worklist. We attempt to 4694 * scatter them about on the list so that they will go off 4695 * at evenly distributed times even if all the filesystems 4696 * are mounted at once. 4697 */ 4698 next += incr; 4699 if (next == 0 || next > syncer_maxdelay) { 4700 start /= 2; 4701 incr /= 2; 4702 if (start == 0) { 4703 start = syncer_maxdelay / 2; 4704 incr = syncer_maxdelay; 4705 } 4706 next = start; 4707 } 4708 bo = &vp->v_bufobj; 4709 BO_LOCK(bo); 4710 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 4711 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 4712 mtx_lock(&sync_mtx); 4713 sync_vnode_count++; 4714 if (mp->mnt_syncer == NULL) { 4715 mp->mnt_syncer = vp; 4716 vp = NULL; 4717 } 4718 mtx_unlock(&sync_mtx); 4719 BO_UNLOCK(bo); 4720 if (vp != NULL) { 4721 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4722 vgone(vp); 4723 vput(vp); 4724 } 4725 } 4726 4727 void 4728 vfs_deallocate_syncvnode(struct mount *mp) 4729 { 4730 struct vnode *vp; 4731 4732 mtx_lock(&sync_mtx); 4733 vp = mp->mnt_syncer; 4734 if (vp != NULL) 4735 mp->mnt_syncer = NULL; 4736 mtx_unlock(&sync_mtx); 4737 if (vp != NULL) 4738 vrele(vp); 4739 } 4740 4741 /* 4742 * Do a lazy sync of the filesystem. 4743 */ 4744 static int 4745 sync_fsync(struct vop_fsync_args *ap) 4746 { 4747 struct vnode *syncvp = ap->a_vp; 4748 struct mount *mp = syncvp->v_mount; 4749 int error, save; 4750 struct bufobj *bo; 4751 4752 /* 4753 * We only need to do something if this is a lazy evaluation. 4754 */ 4755 if (ap->a_waitfor != MNT_LAZY) 4756 return (0); 4757 4758 /* 4759 * Move ourselves to the back of the sync list. 4760 */ 4761 bo = &syncvp->v_bufobj; 4762 BO_LOCK(bo); 4763 vn_syncer_add_to_worklist(bo, syncdelay); 4764 BO_UNLOCK(bo); 4765 4766 /* 4767 * Walk the list of vnodes pushing all that are dirty and 4768 * not already on the sync list. 4769 */ 4770 if (vfs_busy(mp, MBF_NOWAIT) != 0) 4771 return (0); 4772 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 4773 vfs_unbusy(mp); 4774 return (0); 4775 } 4776 save = curthread_pflags_set(TDP_SYNCIO); 4777 /* 4778 * The filesystem at hand may be idle with free vnodes stored in the 4779 * batch. Return them instead of letting them stay there indefinitely. 4780 */ 4781 vnlru_return_batch(mp); 4782 vfs_periodic(mp, MNT_NOWAIT); 4783 error = VFS_SYNC(mp, MNT_LAZY); 4784 curthread_pflags_restore(save); 4785 vn_finished_write(mp); 4786 vfs_unbusy(mp); 4787 return (error); 4788 } 4789 4790 /* 4791 * The syncer vnode is no referenced. 4792 */ 4793 static int 4794 sync_inactive(struct vop_inactive_args *ap) 4795 { 4796 4797 vgone(ap->a_vp); 4798 return (0); 4799 } 4800 4801 /* 4802 * The syncer vnode is no longer needed and is being decommissioned. 4803 * 4804 * Modifications to the worklist must be protected by sync_mtx. 4805 */ 4806 static int 4807 sync_reclaim(struct vop_reclaim_args *ap) 4808 { 4809 struct vnode *vp = ap->a_vp; 4810 struct bufobj *bo; 4811 4812 bo = &vp->v_bufobj; 4813 BO_LOCK(bo); 4814 mtx_lock(&sync_mtx); 4815 if (vp->v_mount->mnt_syncer == vp) 4816 vp->v_mount->mnt_syncer = NULL; 4817 if (bo->bo_flag & BO_ONWORKLST) { 4818 LIST_REMOVE(bo, bo_synclist); 4819 syncer_worklist_len--; 4820 sync_vnode_count--; 4821 bo->bo_flag &= ~BO_ONWORKLST; 4822 } 4823 mtx_unlock(&sync_mtx); 4824 BO_UNLOCK(bo); 4825 4826 return (0); 4827 } 4828 4829 int 4830 vn_need_pageq_flush(struct vnode *vp) 4831 { 4832 struct vm_object *obj; 4833 int need; 4834 4835 MPASS(mtx_owned(VI_MTX(vp))); 4836 need = 0; 4837 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 4838 vm_object_mightbedirty(obj)) 4839 need = 1; 4840 return (need); 4841 } 4842 4843 /* 4844 * Check if vnode represents a disk device 4845 */ 4846 int 4847 vn_isdisk(struct vnode *vp, int *errp) 4848 { 4849 int error; 4850 4851 if (vp->v_type != VCHR) { 4852 error = ENOTBLK; 4853 goto out; 4854 } 4855 error = 0; 4856 dev_lock(); 4857 if (vp->v_rdev == NULL) 4858 error = ENXIO; 4859 else if (vp->v_rdev->si_devsw == NULL) 4860 error = ENXIO; 4861 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 4862 error = ENOTBLK; 4863 dev_unlock(); 4864 out: 4865 if (errp != NULL) 4866 *errp = error; 4867 return (error == 0); 4868 } 4869 4870 /* 4871 * Common filesystem object access control check routine. Accepts a 4872 * vnode's type, "mode", uid and gid, requested access mode, credentials, 4873 * and optional call-by-reference privused argument allowing vaccess() 4874 * to indicate to the caller whether privilege was used to satisfy the 4875 * request (obsoleted). Returns 0 on success, or an errno on failure. 4876 */ 4877 int 4878 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 4879 accmode_t accmode, struct ucred *cred, int *privused) 4880 { 4881 accmode_t dac_granted; 4882 accmode_t priv_granted; 4883 4884 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 4885 ("invalid bit in accmode")); 4886 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 4887 ("VAPPEND without VWRITE")); 4888 4889 /* 4890 * Look for a normal, non-privileged way to access the file/directory 4891 * as requested. If it exists, go with that. 4892 */ 4893 4894 if (privused != NULL) 4895 *privused = 0; 4896 4897 dac_granted = 0; 4898 4899 /* Check the owner. */ 4900 if (cred->cr_uid == file_uid) { 4901 dac_granted |= VADMIN; 4902 if (file_mode & S_IXUSR) 4903 dac_granted |= VEXEC; 4904 if (file_mode & S_IRUSR) 4905 dac_granted |= VREAD; 4906 if (file_mode & S_IWUSR) 4907 dac_granted |= (VWRITE | VAPPEND); 4908 4909 if ((accmode & dac_granted) == accmode) 4910 return (0); 4911 4912 goto privcheck; 4913 } 4914 4915 /* Otherwise, check the groups (first match) */ 4916 if (groupmember(file_gid, cred)) { 4917 if (file_mode & S_IXGRP) 4918 dac_granted |= VEXEC; 4919 if (file_mode & S_IRGRP) 4920 dac_granted |= VREAD; 4921 if (file_mode & S_IWGRP) 4922 dac_granted |= (VWRITE | VAPPEND); 4923 4924 if ((accmode & dac_granted) == accmode) 4925 return (0); 4926 4927 goto privcheck; 4928 } 4929 4930 /* Otherwise, check everyone else. */ 4931 if (file_mode & S_IXOTH) 4932 dac_granted |= VEXEC; 4933 if (file_mode & S_IROTH) 4934 dac_granted |= VREAD; 4935 if (file_mode & S_IWOTH) 4936 dac_granted |= (VWRITE | VAPPEND); 4937 if ((accmode & dac_granted) == accmode) 4938 return (0); 4939 4940 privcheck: 4941 /* 4942 * Build a privilege mask to determine if the set of privileges 4943 * satisfies the requirements when combined with the granted mask 4944 * from above. For each privilege, if the privilege is required, 4945 * bitwise or the request type onto the priv_granted mask. 4946 */ 4947 priv_granted = 0; 4948 4949 if (type == VDIR) { 4950 /* 4951 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 4952 * requests, instead of PRIV_VFS_EXEC. 4953 */ 4954 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 4955 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 4956 priv_granted |= VEXEC; 4957 } else { 4958 /* 4959 * Ensure that at least one execute bit is on. Otherwise, 4960 * a privileged user will always succeed, and we don't want 4961 * this to happen unless the file really is executable. 4962 */ 4963 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 4964 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 4965 !priv_check_cred(cred, PRIV_VFS_EXEC)) 4966 priv_granted |= VEXEC; 4967 } 4968 4969 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 4970 !priv_check_cred(cred, PRIV_VFS_READ)) 4971 priv_granted |= VREAD; 4972 4973 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 4974 !priv_check_cred(cred, PRIV_VFS_WRITE)) 4975 priv_granted |= (VWRITE | VAPPEND); 4976 4977 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 4978 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 4979 priv_granted |= VADMIN; 4980 4981 if ((accmode & (priv_granted | dac_granted)) == accmode) { 4982 /* XXX audit: privilege used */ 4983 if (privused != NULL) 4984 *privused = 1; 4985 return (0); 4986 } 4987 4988 return ((accmode & VADMIN) ? EPERM : EACCES); 4989 } 4990 4991 /* 4992 * Credential check based on process requesting service, and per-attribute 4993 * permissions. 4994 */ 4995 int 4996 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 4997 struct thread *td, accmode_t accmode) 4998 { 4999 5000 /* 5001 * Kernel-invoked always succeeds. 5002 */ 5003 if (cred == NOCRED) 5004 return (0); 5005 5006 /* 5007 * Do not allow privileged processes in jail to directly manipulate 5008 * system attributes. 5009 */ 5010 switch (attrnamespace) { 5011 case EXTATTR_NAMESPACE_SYSTEM: 5012 /* Potentially should be: return (EPERM); */ 5013 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5014 case EXTATTR_NAMESPACE_USER: 5015 return (VOP_ACCESS(vp, accmode, cred, td)); 5016 default: 5017 return (EPERM); 5018 } 5019 } 5020 5021 #ifdef DEBUG_VFS_LOCKS 5022 /* 5023 * This only exists to suppress warnings from unlocked specfs accesses. It is 5024 * no longer ok to have an unlocked VFS. 5025 */ 5026 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL || \ 5027 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 5028 5029 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 5030 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 5031 "Drop into debugger on lock violation"); 5032 5033 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 5034 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 5035 0, "Check for interlock across VOPs"); 5036 5037 int vfs_badlock_print = 1; /* Print lock violations. */ 5038 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 5039 0, "Print lock violations"); 5040 5041 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ 5042 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 5043 0, "Print vnode details on lock violations"); 5044 5045 #ifdef KDB 5046 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 5047 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 5048 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 5049 #endif 5050 5051 static void 5052 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 5053 { 5054 5055 #ifdef KDB 5056 if (vfs_badlock_backtrace) 5057 kdb_backtrace(); 5058 #endif 5059 if (vfs_badlock_vnode) 5060 vn_printf(vp, "vnode "); 5061 if (vfs_badlock_print) 5062 printf("%s: %p %s\n", str, (void *)vp, msg); 5063 if (vfs_badlock_ddb) 5064 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5065 } 5066 5067 void 5068 assert_vi_locked(struct vnode *vp, const char *str) 5069 { 5070 5071 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 5072 vfs_badlock("interlock is not locked but should be", str, vp); 5073 } 5074 5075 void 5076 assert_vi_unlocked(struct vnode *vp, const char *str) 5077 { 5078 5079 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 5080 vfs_badlock("interlock is locked but should not be", str, vp); 5081 } 5082 5083 void 5084 assert_vop_locked(struct vnode *vp, const char *str) 5085 { 5086 int locked; 5087 5088 if (!IGNORE_LOCK(vp)) { 5089 locked = VOP_ISLOCKED(vp); 5090 if (locked == 0 || locked == LK_EXCLOTHER) 5091 vfs_badlock("is not locked but should be", str, vp); 5092 } 5093 } 5094 5095 void 5096 assert_vop_unlocked(struct vnode *vp, const char *str) 5097 { 5098 5099 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 5100 vfs_badlock("is locked but should not be", str, vp); 5101 } 5102 5103 void 5104 assert_vop_elocked(struct vnode *vp, const char *str) 5105 { 5106 5107 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 5108 vfs_badlock("is not exclusive locked but should be", str, vp); 5109 } 5110 #endif /* DEBUG_VFS_LOCKS */ 5111 5112 void 5113 vop_rename_fail(struct vop_rename_args *ap) 5114 { 5115 5116 if (ap->a_tvp != NULL) 5117 vput(ap->a_tvp); 5118 if (ap->a_tdvp == ap->a_tvp) 5119 vrele(ap->a_tdvp); 5120 else 5121 vput(ap->a_tdvp); 5122 vrele(ap->a_fdvp); 5123 vrele(ap->a_fvp); 5124 } 5125 5126 void 5127 vop_rename_pre(void *ap) 5128 { 5129 struct vop_rename_args *a = ap; 5130 5131 #ifdef DEBUG_VFS_LOCKS 5132 if (a->a_tvp) 5133 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5134 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5135 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5136 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5137 5138 /* Check the source (from). */ 5139 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5140 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5141 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5142 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5143 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5144 5145 /* Check the target. */ 5146 if (a->a_tvp) 5147 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5148 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5149 #endif 5150 if (a->a_tdvp != a->a_fdvp) 5151 vhold(a->a_fdvp); 5152 if (a->a_tvp != a->a_fvp) 5153 vhold(a->a_fvp); 5154 vhold(a->a_tdvp); 5155 if (a->a_tvp) 5156 vhold(a->a_tvp); 5157 } 5158 5159 #ifdef DEBUG_VFS_LOCKS 5160 void 5161 vop_strategy_pre(void *ap) 5162 { 5163 struct vop_strategy_args *a; 5164 struct buf *bp; 5165 5166 a = ap; 5167 bp = a->a_bp; 5168 5169 /* 5170 * Cluster ops lock their component buffers but not the IO container. 5171 */ 5172 if ((bp->b_flags & B_CLUSTER) != 0) 5173 return; 5174 5175 if (panicstr == NULL && !BUF_ISLOCKED(bp)) { 5176 if (vfs_badlock_print) 5177 printf( 5178 "VOP_STRATEGY: bp is not locked but should be\n"); 5179 if (vfs_badlock_ddb) 5180 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5181 } 5182 } 5183 5184 void 5185 vop_lock_pre(void *ap) 5186 { 5187 struct vop_lock1_args *a = ap; 5188 5189 if ((a->a_flags & LK_INTERLOCK) == 0) 5190 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5191 else 5192 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 5193 } 5194 5195 void 5196 vop_lock_post(void *ap, int rc) 5197 { 5198 struct vop_lock1_args *a = ap; 5199 5200 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5201 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 5202 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 5203 } 5204 5205 void 5206 vop_unlock_pre(void *ap) 5207 { 5208 struct vop_unlock_args *a = ap; 5209 5210 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 5211 } 5212 5213 void 5214 vop_unlock_post(void *ap, int rc) 5215 { 5216 return; 5217 } 5218 5219 void 5220 vop_need_inactive_pre(void *ap) 5221 { 5222 struct vop_need_inactive_args *a = ap; 5223 5224 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5225 } 5226 5227 void 5228 vop_need_inactive_post(void *ap, int rc) 5229 { 5230 struct vop_need_inactive_args *a = ap; 5231 5232 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5233 } 5234 #endif 5235 5236 void 5237 vop_create_post(void *ap, int rc) 5238 { 5239 struct vop_create_args *a = ap; 5240 5241 if (!rc) 5242 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 5243 } 5244 5245 void 5246 vop_deleteextattr_post(void *ap, int rc) 5247 { 5248 struct vop_deleteextattr_args *a = ap; 5249 5250 if (!rc) 5251 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5252 } 5253 5254 void 5255 vop_link_post(void *ap, int rc) 5256 { 5257 struct vop_link_args *a = ap; 5258 5259 if (!rc) { 5260 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 5261 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 5262 } 5263 } 5264 5265 void 5266 vop_mkdir_post(void *ap, int rc) 5267 { 5268 struct vop_mkdir_args *a = ap; 5269 5270 if (!rc) 5271 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 5272 } 5273 5274 void 5275 vop_mknod_post(void *ap, int rc) 5276 { 5277 struct vop_mknod_args *a = ap; 5278 5279 if (!rc) 5280 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 5281 } 5282 5283 void 5284 vop_reclaim_post(void *ap, int rc) 5285 { 5286 struct vop_reclaim_args *a = ap; 5287 5288 if (!rc) 5289 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE); 5290 } 5291 5292 void 5293 vop_remove_post(void *ap, int rc) 5294 { 5295 struct vop_remove_args *a = ap; 5296 5297 if (!rc) { 5298 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 5299 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 5300 } 5301 } 5302 5303 void 5304 vop_rename_post(void *ap, int rc) 5305 { 5306 struct vop_rename_args *a = ap; 5307 long hint; 5308 5309 if (!rc) { 5310 hint = NOTE_WRITE; 5311 if (a->a_fdvp == a->a_tdvp) { 5312 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 5313 hint |= NOTE_LINK; 5314 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5315 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5316 } else { 5317 hint |= NOTE_EXTEND; 5318 if (a->a_fvp->v_type == VDIR) 5319 hint |= NOTE_LINK; 5320 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5321 5322 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 5323 a->a_tvp->v_type == VDIR) 5324 hint &= ~NOTE_LINK; 5325 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5326 } 5327 5328 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 5329 if (a->a_tvp) 5330 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 5331 } 5332 if (a->a_tdvp != a->a_fdvp) 5333 vdrop(a->a_fdvp); 5334 if (a->a_tvp != a->a_fvp) 5335 vdrop(a->a_fvp); 5336 vdrop(a->a_tdvp); 5337 if (a->a_tvp) 5338 vdrop(a->a_tvp); 5339 } 5340 5341 void 5342 vop_rmdir_post(void *ap, int rc) 5343 { 5344 struct vop_rmdir_args *a = ap; 5345 5346 if (!rc) { 5347 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 5348 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 5349 } 5350 } 5351 5352 void 5353 vop_setattr_post(void *ap, int rc) 5354 { 5355 struct vop_setattr_args *a = ap; 5356 5357 if (!rc) 5358 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5359 } 5360 5361 void 5362 vop_setextattr_post(void *ap, int rc) 5363 { 5364 struct vop_setextattr_args *a = ap; 5365 5366 if (!rc) 5367 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5368 } 5369 5370 void 5371 vop_symlink_post(void *ap, int rc) 5372 { 5373 struct vop_symlink_args *a = ap; 5374 5375 if (!rc) 5376 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 5377 } 5378 5379 void 5380 vop_open_post(void *ap, int rc) 5381 { 5382 struct vop_open_args *a = ap; 5383 5384 if (!rc) 5385 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 5386 } 5387 5388 void 5389 vop_close_post(void *ap, int rc) 5390 { 5391 struct vop_close_args *a = ap; 5392 5393 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 5394 !VN_IS_DOOMED(a->a_vp))) { 5395 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 5396 NOTE_CLOSE_WRITE : NOTE_CLOSE); 5397 } 5398 } 5399 5400 void 5401 vop_read_post(void *ap, int rc) 5402 { 5403 struct vop_read_args *a = ap; 5404 5405 if (!rc) 5406 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 5407 } 5408 5409 void 5410 vop_readdir_post(void *ap, int rc) 5411 { 5412 struct vop_readdir_args *a = ap; 5413 5414 if (!rc) 5415 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 5416 } 5417 5418 static struct knlist fs_knlist; 5419 5420 static void 5421 vfs_event_init(void *arg) 5422 { 5423 knlist_init_mtx(&fs_knlist, NULL); 5424 } 5425 /* XXX - correct order? */ 5426 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 5427 5428 void 5429 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 5430 { 5431 5432 KNOTE_UNLOCKED(&fs_knlist, event); 5433 } 5434 5435 static int filt_fsattach(struct knote *kn); 5436 static void filt_fsdetach(struct knote *kn); 5437 static int filt_fsevent(struct knote *kn, long hint); 5438 5439 struct filterops fs_filtops = { 5440 .f_isfd = 0, 5441 .f_attach = filt_fsattach, 5442 .f_detach = filt_fsdetach, 5443 .f_event = filt_fsevent 5444 }; 5445 5446 static int 5447 filt_fsattach(struct knote *kn) 5448 { 5449 5450 kn->kn_flags |= EV_CLEAR; 5451 knlist_add(&fs_knlist, kn, 0); 5452 return (0); 5453 } 5454 5455 static void 5456 filt_fsdetach(struct knote *kn) 5457 { 5458 5459 knlist_remove(&fs_knlist, kn, 0); 5460 } 5461 5462 static int 5463 filt_fsevent(struct knote *kn, long hint) 5464 { 5465 5466 kn->kn_fflags |= hint; 5467 return (kn->kn_fflags != 0); 5468 } 5469 5470 static int 5471 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 5472 { 5473 struct vfsidctl vc; 5474 int error; 5475 struct mount *mp; 5476 5477 error = SYSCTL_IN(req, &vc, sizeof(vc)); 5478 if (error) 5479 return (error); 5480 if (vc.vc_vers != VFS_CTL_VERS1) 5481 return (EINVAL); 5482 mp = vfs_getvfs(&vc.vc_fsid); 5483 if (mp == NULL) 5484 return (ENOENT); 5485 /* ensure that a specific sysctl goes to the right filesystem. */ 5486 if (strcmp(vc.vc_fstypename, "*") != 0 && 5487 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 5488 vfs_rel(mp); 5489 return (EINVAL); 5490 } 5491 VCTLTOREQ(&vc, req); 5492 error = VFS_SYSCTL(mp, vc.vc_op, req); 5493 vfs_rel(mp); 5494 return (error); 5495 } 5496 5497 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 5498 NULL, 0, sysctl_vfs_ctl, "", 5499 "Sysctl by fsid"); 5500 5501 /* 5502 * Function to initialize a va_filerev field sensibly. 5503 * XXX: Wouldn't a random number make a lot more sense ?? 5504 */ 5505 u_quad_t 5506 init_va_filerev(void) 5507 { 5508 struct bintime bt; 5509 5510 getbinuptime(&bt); 5511 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 5512 } 5513 5514 static int filt_vfsread(struct knote *kn, long hint); 5515 static int filt_vfswrite(struct knote *kn, long hint); 5516 static int filt_vfsvnode(struct knote *kn, long hint); 5517 static void filt_vfsdetach(struct knote *kn); 5518 static struct filterops vfsread_filtops = { 5519 .f_isfd = 1, 5520 .f_detach = filt_vfsdetach, 5521 .f_event = filt_vfsread 5522 }; 5523 static struct filterops vfswrite_filtops = { 5524 .f_isfd = 1, 5525 .f_detach = filt_vfsdetach, 5526 .f_event = filt_vfswrite 5527 }; 5528 static struct filterops vfsvnode_filtops = { 5529 .f_isfd = 1, 5530 .f_detach = filt_vfsdetach, 5531 .f_event = filt_vfsvnode 5532 }; 5533 5534 static void 5535 vfs_knllock(void *arg) 5536 { 5537 struct vnode *vp = arg; 5538 5539 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5540 } 5541 5542 static void 5543 vfs_knlunlock(void *arg) 5544 { 5545 struct vnode *vp = arg; 5546 5547 VOP_UNLOCK(vp); 5548 } 5549 5550 static void 5551 vfs_knl_assert_locked(void *arg) 5552 { 5553 #ifdef DEBUG_VFS_LOCKS 5554 struct vnode *vp = arg; 5555 5556 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 5557 #endif 5558 } 5559 5560 static void 5561 vfs_knl_assert_unlocked(void *arg) 5562 { 5563 #ifdef DEBUG_VFS_LOCKS 5564 struct vnode *vp = arg; 5565 5566 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 5567 #endif 5568 } 5569 5570 int 5571 vfs_kqfilter(struct vop_kqfilter_args *ap) 5572 { 5573 struct vnode *vp = ap->a_vp; 5574 struct knote *kn = ap->a_kn; 5575 struct knlist *knl; 5576 5577 switch (kn->kn_filter) { 5578 case EVFILT_READ: 5579 kn->kn_fop = &vfsread_filtops; 5580 break; 5581 case EVFILT_WRITE: 5582 kn->kn_fop = &vfswrite_filtops; 5583 break; 5584 case EVFILT_VNODE: 5585 kn->kn_fop = &vfsvnode_filtops; 5586 break; 5587 default: 5588 return (EINVAL); 5589 } 5590 5591 kn->kn_hook = (caddr_t)vp; 5592 5593 v_addpollinfo(vp); 5594 if (vp->v_pollinfo == NULL) 5595 return (ENOMEM); 5596 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 5597 vhold(vp); 5598 knlist_add(knl, kn, 0); 5599 5600 return (0); 5601 } 5602 5603 /* 5604 * Detach knote from vnode 5605 */ 5606 static void 5607 filt_vfsdetach(struct knote *kn) 5608 { 5609 struct vnode *vp = (struct vnode *)kn->kn_hook; 5610 5611 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 5612 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 5613 vdrop(vp); 5614 } 5615 5616 /*ARGSUSED*/ 5617 static int 5618 filt_vfsread(struct knote *kn, long hint) 5619 { 5620 struct vnode *vp = (struct vnode *)kn->kn_hook; 5621 struct vattr va; 5622 int res; 5623 5624 /* 5625 * filesystem is gone, so set the EOF flag and schedule 5626 * the knote for deletion. 5627 */ 5628 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 5629 VI_LOCK(vp); 5630 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 5631 VI_UNLOCK(vp); 5632 return (1); 5633 } 5634 5635 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 5636 return (0); 5637 5638 VI_LOCK(vp); 5639 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 5640 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 5641 VI_UNLOCK(vp); 5642 return (res); 5643 } 5644 5645 /*ARGSUSED*/ 5646 static int 5647 filt_vfswrite(struct knote *kn, long hint) 5648 { 5649 struct vnode *vp = (struct vnode *)kn->kn_hook; 5650 5651 VI_LOCK(vp); 5652 5653 /* 5654 * filesystem is gone, so set the EOF flag and schedule 5655 * the knote for deletion. 5656 */ 5657 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 5658 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 5659 5660 kn->kn_data = 0; 5661 VI_UNLOCK(vp); 5662 return (1); 5663 } 5664 5665 static int 5666 filt_vfsvnode(struct knote *kn, long hint) 5667 { 5668 struct vnode *vp = (struct vnode *)kn->kn_hook; 5669 int res; 5670 5671 VI_LOCK(vp); 5672 if (kn->kn_sfflags & hint) 5673 kn->kn_fflags |= hint; 5674 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 5675 kn->kn_flags |= EV_EOF; 5676 VI_UNLOCK(vp); 5677 return (1); 5678 } 5679 res = (kn->kn_fflags != 0); 5680 VI_UNLOCK(vp); 5681 return (res); 5682 } 5683 5684 /* 5685 * Returns whether the directory is empty or not. 5686 * If it is empty, the return value is 0; otherwise 5687 * the return value is an error value (which may 5688 * be ENOTEMPTY). 5689 */ 5690 int 5691 vfs_emptydir(struct vnode *vp) 5692 { 5693 struct uio uio; 5694 struct iovec iov; 5695 struct dirent *dirent, *dp, *endp; 5696 int error, eof; 5697 5698 error = 0; 5699 eof = 0; 5700 5701 ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); 5702 5703 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK); 5704 iov.iov_base = dirent; 5705 iov.iov_len = sizeof(struct dirent); 5706 5707 uio.uio_iov = &iov; 5708 uio.uio_iovcnt = 1; 5709 uio.uio_offset = 0; 5710 uio.uio_resid = sizeof(struct dirent); 5711 uio.uio_segflg = UIO_SYSSPACE; 5712 uio.uio_rw = UIO_READ; 5713 uio.uio_td = curthread; 5714 5715 while (eof == 0 && error == 0) { 5716 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof, 5717 NULL, NULL); 5718 if (error != 0) 5719 break; 5720 endp = (void *)((uint8_t *)dirent + 5721 sizeof(struct dirent) - uio.uio_resid); 5722 for (dp = dirent; dp < endp; 5723 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) { 5724 if (dp->d_type == DT_WHT) 5725 continue; 5726 if (dp->d_namlen == 0) 5727 continue; 5728 if (dp->d_type != DT_DIR && 5729 dp->d_type != DT_UNKNOWN) { 5730 error = ENOTEMPTY; 5731 break; 5732 } 5733 if (dp->d_namlen > 2) { 5734 error = ENOTEMPTY; 5735 break; 5736 } 5737 if (dp->d_namlen == 1 && 5738 dp->d_name[0] != '.') { 5739 error = ENOTEMPTY; 5740 break; 5741 } 5742 if (dp->d_namlen == 2 && 5743 dp->d_name[1] != '.') { 5744 error = ENOTEMPTY; 5745 break; 5746 } 5747 uio.uio_resid = sizeof(struct dirent); 5748 } 5749 } 5750 free(dirent, M_TEMP); 5751 return (error); 5752 } 5753 5754 int 5755 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 5756 { 5757 int error; 5758 5759 if (dp->d_reclen > ap->a_uio->uio_resid) 5760 return (ENAMETOOLONG); 5761 error = uiomove(dp, dp->d_reclen, ap->a_uio); 5762 if (error) { 5763 if (ap->a_ncookies != NULL) { 5764 if (ap->a_cookies != NULL) 5765 free(ap->a_cookies, M_TEMP); 5766 ap->a_cookies = NULL; 5767 *ap->a_ncookies = 0; 5768 } 5769 return (error); 5770 } 5771 if (ap->a_ncookies == NULL) 5772 return (0); 5773 5774 KASSERT(ap->a_cookies, 5775 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 5776 5777 *ap->a_cookies = realloc(*ap->a_cookies, 5778 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 5779 (*ap->a_cookies)[*ap->a_ncookies] = off; 5780 *ap->a_ncookies += 1; 5781 return (0); 5782 } 5783 5784 /* 5785 * Mark for update the access time of the file if the filesystem 5786 * supports VOP_MARKATIME. This functionality is used by execve and 5787 * mmap, so we want to avoid the I/O implied by directly setting 5788 * va_atime for the sake of efficiency. 5789 */ 5790 void 5791 vfs_mark_atime(struct vnode *vp, struct ucred *cred) 5792 { 5793 struct mount *mp; 5794 5795 mp = vp->v_mount; 5796 ASSERT_VOP_LOCKED(vp, "vfs_mark_atime"); 5797 if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) 5798 (void)VOP_MARKATIME(vp); 5799 } 5800 5801 /* 5802 * The purpose of this routine is to remove granularity from accmode_t, 5803 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 5804 * VADMIN and VAPPEND. 5805 * 5806 * If it returns 0, the caller is supposed to continue with the usual 5807 * access checks using 'accmode' as modified by this routine. If it 5808 * returns nonzero value, the caller is supposed to return that value 5809 * as errno. 5810 * 5811 * Note that after this routine runs, accmode may be zero. 5812 */ 5813 int 5814 vfs_unixify_accmode(accmode_t *accmode) 5815 { 5816 /* 5817 * There is no way to specify explicit "deny" rule using 5818 * file mode or POSIX.1e ACLs. 5819 */ 5820 if (*accmode & VEXPLICIT_DENY) { 5821 *accmode = 0; 5822 return (0); 5823 } 5824 5825 /* 5826 * None of these can be translated into usual access bits. 5827 * Also, the common case for NFSv4 ACLs is to not contain 5828 * either of these bits. Caller should check for VWRITE 5829 * on the containing directory instead. 5830 */ 5831 if (*accmode & (VDELETE_CHILD | VDELETE)) 5832 return (EPERM); 5833 5834 if (*accmode & VADMIN_PERMS) { 5835 *accmode &= ~VADMIN_PERMS; 5836 *accmode |= VADMIN; 5837 } 5838 5839 /* 5840 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 5841 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 5842 */ 5843 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 5844 5845 return (0); 5846 } 5847 5848 /* 5849 * Clear out a doomed vnode (if any) and replace it with a new one as long 5850 * as the fs is not being unmounted. Return the root vnode to the caller. 5851 */ 5852 static int __noinline 5853 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 5854 { 5855 struct vnode *vp; 5856 int error; 5857 5858 restart: 5859 if (mp->mnt_rootvnode != NULL) { 5860 MNT_ILOCK(mp); 5861 vp = mp->mnt_rootvnode; 5862 if (vp != NULL) { 5863 if (!VN_IS_DOOMED(vp)) { 5864 vrefact(vp); 5865 MNT_IUNLOCK(mp); 5866 error = vn_lock(vp, flags); 5867 if (error == 0) { 5868 *vpp = vp; 5869 return (0); 5870 } 5871 vrele(vp); 5872 goto restart; 5873 } 5874 /* 5875 * Clear the old one. 5876 */ 5877 mp->mnt_rootvnode = NULL; 5878 } 5879 MNT_IUNLOCK(mp); 5880 if (vp != NULL) { 5881 /* 5882 * Paired with a fence in vfs_op_thread_exit(). 5883 */ 5884 atomic_thread_fence_acq(); 5885 vfs_op_barrier_wait(mp); 5886 vrele(vp); 5887 } 5888 } 5889 error = VFS_CACHEDROOT(mp, flags, vpp); 5890 if (error != 0) 5891 return (error); 5892 if (mp->mnt_vfs_ops == 0) { 5893 MNT_ILOCK(mp); 5894 if (mp->mnt_vfs_ops != 0) { 5895 MNT_IUNLOCK(mp); 5896 return (0); 5897 } 5898 if (mp->mnt_rootvnode == NULL) { 5899 vrefact(*vpp); 5900 mp->mnt_rootvnode = *vpp; 5901 } else { 5902 if (mp->mnt_rootvnode != *vpp) { 5903 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 5904 panic("%s: mismatch between vnode returned " 5905 " by VFS_CACHEDROOT and the one cached " 5906 " (%p != %p)", 5907 __func__, *vpp, mp->mnt_rootvnode); 5908 } 5909 } 5910 } 5911 MNT_IUNLOCK(mp); 5912 } 5913 return (0); 5914 } 5915 5916 int 5917 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 5918 { 5919 struct vnode *vp; 5920 int error; 5921 5922 if (!vfs_op_thread_enter(mp)) 5923 return (vfs_cache_root_fallback(mp, flags, vpp)); 5924 vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode); 5925 if (vp == NULL || VN_IS_DOOMED(vp)) { 5926 vfs_op_thread_exit(mp); 5927 return (vfs_cache_root_fallback(mp, flags, vpp)); 5928 } 5929 vrefact(vp); 5930 vfs_op_thread_exit(mp); 5931 error = vn_lock(vp, flags); 5932 if (error != 0) { 5933 vrele(vp); 5934 return (vfs_cache_root_fallback(mp, flags, vpp)); 5935 } 5936 *vpp = vp; 5937 return (0); 5938 } 5939 5940 struct vnode * 5941 vfs_cache_root_clear(struct mount *mp) 5942 { 5943 struct vnode *vp; 5944 5945 /* 5946 * ops > 0 guarantees there is nobody who can see this vnode 5947 */ 5948 MPASS(mp->mnt_vfs_ops > 0); 5949 vp = mp->mnt_rootvnode; 5950 mp->mnt_rootvnode = NULL; 5951 return (vp); 5952 } 5953 5954 void 5955 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 5956 { 5957 5958 MPASS(mp->mnt_vfs_ops > 0); 5959 vrefact(vp); 5960 mp->mnt_rootvnode = vp; 5961 } 5962 5963 /* 5964 * These are helper functions for filesystems to traverse all 5965 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 5966 * 5967 * This interface replaces MNT_VNODE_FOREACH. 5968 */ 5969 5970 5971 struct vnode * 5972 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 5973 { 5974 struct vnode *vp; 5975 5976 if (should_yield()) 5977 kern_yield(PRI_USER); 5978 MNT_ILOCK(mp); 5979 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5980 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 5981 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 5982 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 5983 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 5984 continue; 5985 VI_LOCK(vp); 5986 if (VN_IS_DOOMED(vp)) { 5987 VI_UNLOCK(vp); 5988 continue; 5989 } 5990 break; 5991 } 5992 if (vp == NULL) { 5993 __mnt_vnode_markerfree_all(mvp, mp); 5994 /* MNT_IUNLOCK(mp); -- done in above function */ 5995 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 5996 return (NULL); 5997 } 5998 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 5999 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6000 MNT_IUNLOCK(mp); 6001 return (vp); 6002 } 6003 6004 struct vnode * 6005 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 6006 { 6007 struct vnode *vp; 6008 6009 *mvp = vn_alloc_marker(mp); 6010 MNT_ILOCK(mp); 6011 MNT_REF(mp); 6012 6013 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 6014 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6015 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6016 continue; 6017 VI_LOCK(vp); 6018 if (VN_IS_DOOMED(vp)) { 6019 VI_UNLOCK(vp); 6020 continue; 6021 } 6022 break; 6023 } 6024 if (vp == NULL) { 6025 MNT_REL(mp); 6026 MNT_IUNLOCK(mp); 6027 vn_free_marker(*mvp); 6028 *mvp = NULL; 6029 return (NULL); 6030 } 6031 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6032 MNT_IUNLOCK(mp); 6033 return (vp); 6034 } 6035 6036 void 6037 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 6038 { 6039 6040 if (*mvp == NULL) { 6041 MNT_IUNLOCK(mp); 6042 return; 6043 } 6044 6045 mtx_assert(MNT_MTX(mp), MA_OWNED); 6046 6047 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6048 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6049 MNT_REL(mp); 6050 MNT_IUNLOCK(mp); 6051 vn_free_marker(*mvp); 6052 *mvp = NULL; 6053 } 6054 6055 /* 6056 * These are helper functions for filesystems to traverse their 6057 * active vnodes. See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h 6058 */ 6059 static void 6060 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 6061 { 6062 6063 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6064 6065 MNT_ILOCK(mp); 6066 MNT_REL(mp); 6067 MNT_IUNLOCK(mp); 6068 vn_free_marker(*mvp); 6069 *mvp = NULL; 6070 } 6071 6072 /* 6073 * Relock the mp mount vnode list lock with the vp vnode interlock in the 6074 * conventional lock order during mnt_vnode_next_active iteration. 6075 * 6076 * On entry, the mount vnode list lock is held and the vnode interlock is not. 6077 * The list lock is dropped and reacquired. On success, both locks are held. 6078 * On failure, the mount vnode list lock is held but the vnode interlock is 6079 * not, and the procedure may have yielded. 6080 */ 6081 static bool 6082 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp, 6083 struct vnode *vp) 6084 { 6085 const struct vnode *tmp; 6086 bool held, ret; 6087 6088 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 6089 TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp, 6090 ("%s: bad marker", __func__)); 6091 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 6092 ("%s: inappropriate vnode", __func__)); 6093 ASSERT_VI_UNLOCKED(vp, __func__); 6094 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6095 6096 ret = false; 6097 6098 TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist); 6099 TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist); 6100 6101 /* 6102 * Use a hold to prevent vp from disappearing while the mount vnode 6103 * list lock is dropped and reacquired. Normally a hold would be 6104 * acquired with vhold(), but that might try to acquire the vnode 6105 * interlock, which would be a LOR with the mount vnode list lock. 6106 */ 6107 held = refcount_acquire_if_not_zero(&vp->v_holdcnt); 6108 mtx_unlock(&mp->mnt_listmtx); 6109 if (!held) 6110 goto abort; 6111 VI_LOCK(vp); 6112 if (!refcount_release_if_not_last(&vp->v_holdcnt)) { 6113 vdropl(vp); 6114 goto abort; 6115 } 6116 mtx_lock(&mp->mnt_listmtx); 6117 6118 /* 6119 * Determine whether the vnode is still the next one after the marker, 6120 * excepting any other markers. If the vnode has not been doomed by 6121 * vgone() then the hold should have ensured that it remained on the 6122 * active list. If it has been doomed but is still on the active list, 6123 * don't abort, but rather skip over it (avoid spinning on doomed 6124 * vnodes). 6125 */ 6126 tmp = mvp; 6127 do { 6128 tmp = TAILQ_NEXT(tmp, v_actfreelist); 6129 } while (tmp != NULL && tmp->v_type == VMARKER); 6130 if (tmp != vp) { 6131 mtx_unlock(&mp->mnt_listmtx); 6132 VI_UNLOCK(vp); 6133 goto abort; 6134 } 6135 6136 ret = true; 6137 goto out; 6138 abort: 6139 maybe_yield(); 6140 mtx_lock(&mp->mnt_listmtx); 6141 out: 6142 if (ret) 6143 ASSERT_VI_LOCKED(vp, __func__); 6144 else 6145 ASSERT_VI_UNLOCKED(vp, __func__); 6146 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6147 return (ret); 6148 } 6149 6150 static struct vnode * 6151 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 6152 { 6153 struct vnode *vp, *nvp; 6154 6155 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6156 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6157 restart: 6158 vp = TAILQ_NEXT(*mvp, v_actfreelist); 6159 while (vp != NULL) { 6160 if (vp->v_type == VMARKER) { 6161 vp = TAILQ_NEXT(vp, v_actfreelist); 6162 continue; 6163 } 6164 /* 6165 * Try-lock because this is the wrong lock order. If that does 6166 * not succeed, drop the mount vnode list lock and try to 6167 * reacquire it and the vnode interlock in the right order. 6168 */ 6169 if (!VI_TRYLOCK(vp) && 6170 !mnt_vnode_next_active_relock(*mvp, mp, vp)) 6171 goto restart; 6172 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 6173 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 6174 ("alien vnode on the active list %p %p", vp, mp)); 6175 if (vp->v_mount == mp && !VN_IS_DOOMED(vp)) 6176 break; 6177 nvp = TAILQ_NEXT(vp, v_actfreelist); 6178 VI_UNLOCK(vp); 6179 vp = nvp; 6180 } 6181 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 6182 6183 /* Check if we are done */ 6184 if (vp == NULL) { 6185 mtx_unlock(&mp->mnt_listmtx); 6186 mnt_vnode_markerfree_active(mvp, mp); 6187 return (NULL); 6188 } 6189 TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist); 6190 mtx_unlock(&mp->mnt_listmtx); 6191 ASSERT_VI_LOCKED(vp, "active iter"); 6192 KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp)); 6193 return (vp); 6194 } 6195 6196 struct vnode * 6197 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 6198 { 6199 6200 if (should_yield()) 6201 kern_yield(PRI_USER); 6202 mtx_lock(&mp->mnt_listmtx); 6203 return (mnt_vnode_next_active(mvp, mp)); 6204 } 6205 6206 struct vnode * 6207 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp) 6208 { 6209 struct vnode *vp; 6210 6211 *mvp = vn_alloc_marker(mp); 6212 MNT_ILOCK(mp); 6213 MNT_REF(mp); 6214 MNT_IUNLOCK(mp); 6215 6216 mtx_lock(&mp->mnt_listmtx); 6217 vp = TAILQ_FIRST(&mp->mnt_activevnodelist); 6218 if (vp == NULL) { 6219 mtx_unlock(&mp->mnt_listmtx); 6220 mnt_vnode_markerfree_active(mvp, mp); 6221 return (NULL); 6222 } 6223 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist); 6224 return (mnt_vnode_next_active(mvp, mp)); 6225 } 6226 6227 void 6228 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 6229 { 6230 6231 if (*mvp == NULL) 6232 return; 6233 6234 mtx_lock(&mp->mnt_listmtx); 6235 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 6236 mtx_unlock(&mp->mnt_listmtx); 6237 mnt_vnode_markerfree_active(mvp, mp); 6238 } 6239